When AI Agents Go Off Script: What the OpenAI and Hugging Face Incident Means for Defenders
As AI agents become more autonomous, understanding behavior is becoming even more important. The OpenAI and Hugging Face incident highlights why behavioral security is foundational to securely deploying AI in the enterprise.
No items found.
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.
No items found.
Share
22
Jul 2026
The recent security incident disclosed jointly by OpenAI and Hugging Face is quickly becoming one of the most discussed AI security stories of the year; it's one that has generated great discussion amongst our researchers and technical teams here at Darktrace.
During an internal evaluation of advanced OpenAI models, an autonomous agent reportedly escaped its intended testing boundaries, obtained internet access, and compromised Hugging Face infrastructure while attempting to evaluate and benchmark the model's cyber capabilities. OpenAI has described it as an unprecedented cyber incident and is collaborating with Hugging Face to investigate and share its findings publicly to help defenders understand the implications and improve AI safety across the community.
What this means for organizations deploying AI agents
The most important takeaway for security leaders is not that the AI model is capable of complex and effective chained attacks. It's what this tells us about the guardrails and protections that need to be put in place to ensure these increasingly capable models are safely used within our businesses.
According to OpenAI's investigation, the models were not acting with malicious intent. They were attempting to achieve a legitimate assigned objective. In pursuing that objective, they identified a path that involved obtaining information from a third party's environment and followed it to completion. From the model's perspective, that appears to have been an effective solution to the problem it was given.
Many organizations still hold the assumption that giving an AI agent a legitimate goal will naturally result in legitimate behavior. Increasingly, that is not something defenders can safely assume.
A human who encounters a "permission denied" message generally understands the technical restriction as well as the social meaning behind it. An agent may interpret that message only as evidence that one method to achieve its objective has failed. What makes this incident notable is the agent’s ability to keep pursuing the objective, revise its plan, and combine many individually useful actions into an outcome its designers did not intend.
This is not an isolated challenge. Recent research from the UK's AI Security Institute found that frontier models often attempt to "cheat" evaluations by finding unintended shortcuts to success. More interestingly, that behavior is not always reflected in chain-of-thought outputs and is not always admitted when models are questioned about their actions. One lesson for defenders is that understanding what an AI system says it is doing is not always the same as understanding its behavior and what it is actually doing.
Why behavioral security is an essential security foundation
That's the principle behind Darktrace / SECURE AI and our broader approach to behavioral security. Understanding intent or putting in place static guardrails is not enough; security teams need visibility into how agents behave over time, what resources they interact with, how their behavior changes, and whether they begin pursuing objectives in unexpected ways. Research released by NIST earlier this summer reinforced that rules alone are not sufficient to defend against AI threats, with one conclusion being that "there will always be a way to prompt an AI system to disregard its rules - it's just a matter of finding it."
Another interesting lesson is that nothing described in OpenAI's preliminary findings appears fundamentally impossible for a skilled human attacker. But it does highlight the accessibility and scale that could be unlocked for attackers as increasingly capable models become more widely available. Resource constraints that attackers deal with will decrease as frontier models can help execute elements of those attack workflows autonomously, persist through failure, and iterate towards an objective. Capabilities that were once limited to skilled attackers can increasingly be supplemented with these systems, lowering the barrier to entry to conduct complex cyber operations.
It is incredibly positive to see OpenAI and Hugging Face investigating this incident collaboratively and sharing their findings publicly. Transparency and collaboration around incidents like these helps the entire security community better understand how rapidly AI capabilities are evolving and where we need to improve safeguards. It also reinforces why Darktrace is excited to be partnering with OpenAI through the Daybreak Cyber Partner Program: to explore how advanced AI can be used safely and responsibly in support of defenders while developing the visibility, governance, and security controls needed for increasingly autonomous systems.
This incident is an important reminder that capable AI systems will not always behave in the ways we expect. As organizations deploy AI agents across critical workflows, understanding behavior is an even more important foundation for effective security.
[related-resource]
Secure Your AI Adoption
Practical guidance for securing LLM deployments, AI-integrated workflows, and agent-based systems, including controls to manage AI risk.
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。
