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
Tyler Fornes
Senior Security Analyst at Expel
Share
11
Mar 2019
The following guest-authored blog post examines an advanced cyber-threat discovered by Darktrace on a customer’s network.
At Expel — a managed security provider — our analysts get to use a lot of really cool technologies every day, including Darktrace. Given its popularity among our customers, we thought it would be useful to demonstrate how Darktrace helps us identify and triage potential security threats.
Take a look at this alert. It was triggered via a violation of one of the pre-packaged model breaches for Device / AT Service Scheduled Task.
To triage this specific alert, we need to answer the following questions:
What were the triggers that caused the model to alert?
Which host was the Scheduled Task created on?
Were any files transferred?
Is this activity commonly seen between these hosts?
By answering these questions, we can determine whether or not this alert is related to malicious activity. First, we need to gather additional evidence using the Darktrace console.
At this point, we know the model breach Device / AT Service Scheduled Task was triggered. But what does that mean? Let’s view the model and explore the logic.
Looking at the logic behind this model breach, we see that any message containing the strings “atsvc” and “IPC$” will match this model breach. And because the frequency has been set to “> 0 in 60 mins,” we can assume that once this activity is seen just one time, it’ll trigger an alert. By understanding this logic, we now know:
Next, let’s grab some data. We opened the Model Breach Event Log to see the related events observed for this model breach. There was a successful DCE-RPC bind, followed by SMB Write/Read success containing the keywords “atsvc” and “IPC$.”
We turned to the View advanced search for this event feature of the Model Breach Event Log for even more info.
The advanced search results for this model breach revealed two distinct messages. There’s a successful NTLM authentication message for the account “appadmin.” Since NTLM is commonly used with SMB for authentication, this is likely the account being used by the source machine to establish the SMB session.
Immediately after this authentication, we see the following DCE-RPC message for a named pipe being created involving atsvc:
We see that the RPC bind was created referencing the SASec interface. Based on a quick online search, we learned that the SASec interface “only includes methods for manipulating account information, because most SASec-created task configuration is stored in the file system using the .JOB file format0.”
One possible explanation for this connection is that it was made to query information about a scheduled task defined within the .JOB format, rather than a new scheduled task being created on the host. However, Darktrace doesn’t show any messages mentioning a file with the extension “.JOB” within this model breach. So we kept digging for answers.
By querying “*.JOB AND SMB” within the timeframe of the activity we’ve already observed, some promising results appeared:
We observed three unique .JOB files being accessed over SMB during the exact time of our previous observations. Considering the hosts and the timeframe, we correlated this activity to the original model breach.
So we know the following:
To answer the last investigative question, we used the query “AV.job AND SMB” over the past 60 days. This query returned daily entries for identical activity dating back several months. The activity occurred around the same time each day, involving the same hosts and file paths.
This was starting to smell like legitimate activity, but we still wanted to analyze the contents of the requested file AV.job. We created a packet capture for a five-minute window around the timeframe of the source IP address observed in the model breach.
Once we collected the PCAP, we downloaded and analyzed it in Wireshark, and then extracted the transferred files using the Export Objects feature.
The contents of this file refer to an executable in the location C:\Program Files\Sophos\Sophos Anti-Virus\BackgroundScanClient.exe. Judging by the name of the .JOB file this was found in, it was likely a legitimate scheduled task created to perform an antivirus scan on the endpoint each morning.
Reviewing our original analysis questions, we could confidently answer all four questions:
Darktrace’s cyber defense platform allowed our analysts to quickly confirm and scope potential threat activity and identify network-based indicators (NBIs) related to an attack. It can also generate additional, host-based indicators (HBIs) to supplement your investigation. In summary, Darktrace AI enables our Expel analysts quickly and efficiently scope an incident or hunt for threats across the entire organization — without the need for exhaustive data collection and offline parsing by an analyst.
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.
“I wonder if running that tool influenced the connectivity in some way.”
“I might need to target down or reboot, but first, I'll investigate the network using ARP. I wonder if ARP is responding properly. If I see "ip neigh stale," that could indicate a problem. I think arping might help in this situation, but I’ll need to keep an eye on how the network is behaving. It's all about figuring out where the issue lies!”
図10：エージェントの異常なLDAPアクティビティに対してトリガーされたこのLinuxサーバーに対するモデル検知は、Cyber AI Analystによって自律的に調査され、その調査結果にはエージェントの異常なSMBセッションアクティビティに関する内容が含まれ、Cyber AIAnalystインシデントに記録されました。
DarktraceのCyber AI Analystは、エージェントのアクティビティによってトリガーされたモデル検知を自律的に調査し、その調査結果を1つの重要なインシデントレポートにまとめました。インシデント報告書の内容は、エージェントによるネットワークスキャンから始まりコーディング演習内容のSMBベースの上書きで終わる、エージェントの行動の各段階を詳しくカバーしています。
図20：Cyber AI Analystは、モデル検知に対する自律的な調査を通じて得られた所見をつなぎ合わせ、エージェントの進行状況のタイムラインを作成し、アクティビティを視覚的に表したグラフとともに提供しました。
Cyber Verification Program下のAnthropic Opus 4.6モデルなど、他のフロンティアモデルを使用したエージェントも、Darktraceのシミュレートされた環境で同じ設定とタスクを使ってテストされました。反復の度に、エージェントは演習の内容または採点プログラムを変更して不正を行おうとしました。基盤となるモデルに関係なく、エージェントは自分が置かれた環境を侵害することで、100％のスコアを達成する方法を見つけることに成功していました。これらのエージェントによるハッキングは、アクティビティの内容こそ違っていましたが、すべてその環境のベースラインの振る舞いから大きく逸脱した動作を引き起こしていました。これらの逸脱はDarktraceプラットフォームに十分な検知の機会を提供し、エージェントの逃げ場はありませんでした。
エージェンティックハーネスは会話履歴をローカルに保存しますが、ダークトレースの研究者たちは、保存されたAIの応答が本当にモデルによって生成されたものであるかどうかの検証が行われていないことを発見しました。また、この設計上の選択がAnthropic Claude Code、AWS Kiro-CLI、OpenAI Codex、およびオープンソースのPiにわたって共通していることを確認しました。
会話履歴は多くの場合クライアント側に保存されます。例えば、Anthropic Claude Code、OpenAI Codex、AWS Kiro-CLI、Piなどでもそうです。そのため、ユーザーはセッションを自由に再開でき、一部の製品にはその履歴を操作する機能が組み込まれています。たとえば、会話の中の特定の時点に戻り、送信したメッセージを編集して、会話を別の軌道へ進めていくことができます。きわめて重要な点として、私たちが調査したすべてのケースにおいて、保存されているAIの応答が対応するモデルから生成されたものであり改ざんされていないことの検証は行われていませんでした。
