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
Max Heinemeyer
Global Field CISO
Share
Introduction: Mirai Malware attack on IoT devices
The rise of Internet of Things (IoT) devices, like Internet-connected cameras used in CCTV surveillance, has introduced new vulnerabilities to both personal and business environments.
With attackers exploiting the inadequate security measures typically associated with IoT devices, traditional antivirus and legacy security solutions fall short. To address this critical gap, Darktrace’s Cyber AI Platform provides advanced protection for these increasingly essential yet vulnerable technologies.
Attack Overview: Mirai Malware
In late May, Darktrace detected the Mirai malware infecting an internet-facing DVR camera owned by a logistics company in Canada.
Mirai, first discovered in 2016, continuously scans the Internet for the IP addresses of vulnerable devices in the Internet of Things (IoT), and then turns these devices into bots that can be used as part of botnets for large-scale network attacks. These attacks are often difficult to detect, as IoT devices seamlessly integrate into digital infrastructure, creating a vastly expanded attack surface for organizations.
By drawing on a bespoke, evolving understanding of what is normal for the network, Darktrace caught each stage in this attack’s lifecycle. However, because this company was still conducting their 30-day Proof of Value, Darktrace's Autonomous Response was not in active and the attack continued past the point of initial compromise. Had Darktrace's Autonomous Response been in active mode, the attack would not have advanced past initial compromise.
Attack Timeline
Figure 1: This timeline roughly outlines the major attack phases over three days of activity
Technical analysis
At the time of the initial breach, this specific botnet’s infrastructure was not yet known to open source intelligence (OSINT). Darktrace, however, detected an EXE download from a location not previously visited by the network.
After the first anomalous EXE download, another was downloaded approximately twenty minutes later. The malware then reached out to multiple IP addresses that were statistically rare for the network. Specifically, the compromised device began transferring large amounts of data to an IP address in China.
Figure 2: An overview of Darktrace detections
Darktrace, by leveraging machine learning algorithms in a protocol agnostic capacity, analyzed this individual device’s transfers within the context of a continuously evolving understanding of what is normal both for this device and for the wider organization. It was therefore able to immediately flag all of these transfers as unusual.
This activity was fully investigated and reported on by Darktrace’s Cyber AI Analyst. A sample of the AI Analyst’s report is shown below. The Suspicious File Download, the Unusual Repeated Connections, and the Unusual External Data Transfer are all presented as unexpected events that call for further investigation. The destination IP of the suspicious download was determined to have 100% hostname rarity relative to what is normal for the organization.
Figure 3: Darktrace’s Cyber AI Analyst autonomously triages the attack
Moreover, the hash of the file, highlighted in a red box in the figure above, revealed that it was a well-known file related to the Mirai Botnet. However, with no antivirus or other security defending the IoT camera, this had gone undetected.
A one-click analysis of the infected device shows a timeline of the model breaches that occurred and graphs the activity to give the report’s readers a quick understanding of the successive stages of the attack. Here, we see the second and third stages of the attack’s lifecycle, in which it starts DDoS against other devices in order to complete its mission while simultaneously continuing outgoing connections to rare destinations in order to sustain its presence.
Figure 4: The device event log showing the list of model breaches on May 23
Conclusion
Interestingly, the client saw no indicators of this activity beyond a sluggish network. This change in network activity was only explained after being identified by Darktrace. Once the client was promptly notified, the compromise was deescalated, and discovering it was a DVR security camera, the client took the device offline.
As this customer was still concluding their trial deployment, Darktrace was not in full autonomous mode. However, if it had been, Darktrace would have responded with a two-tiered action to prevent the device from communicating with the malicious endpoint, cutting the compromised connection before the attack had gained its foothold.
Darktrace model breaches:
Anomalous Connection / Uncommon 1GiB Outbound
Unusual Activity / Unusual External Activity
Unusual Activity / Enhanced Unusual External Data Transfer
Unusual Activity / Unusual External Data to New IPs
Device / Initial Breach Chain Compromise
Anomalous Server Activity / Outgoing from Server
Anomalous Connection / Data Sent to New External Device
Anomalous Connection / Multiple Connections to New External UDP Port
Anomalous Connection / Data Sent to Rare Domain
Anomalous File / EXE from Rare External Location
Anomalous File / Internet Facing System File Download
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。
