Anthropic security incidents and Claude misuse reports
Anthropic · San Francisco, CA, USA · founded 2021 · Claude, Claude Code, Claude for Enterprise
This is a sourced record of publicly reported security and safety incidents involving Anthropic and its Claude models. The entries so far are Anthropic's own disclosures: a 2025 report that it detected and disrupted a state-linked group using Claude to automate much of a cyber-espionage operation, and its 2024 research on many-shot jailbreaking.
Anthropic's notable disclosures are largely self-published safety research and misuse reports — it has proactively documented attempts to weaponise Claude and studied jailbreak techniques, positioning disclosure as part of its safety posture.
Reported incidents (2)
Anthropic published a report stating it had detected and disrupted a state-linked group that used Claude (including agentic coding tooling) to automate much of a cyber-espionage operation against a set of targets. The disclosure was Anthropic's own, framed as evidence that agentic models lower the barrier to automated attacks.
Anthropic documented "many-shot jailbreaking", a technique that uses a long context of staged examples to erode a model's refusals, and disclosed it along with mitigations — an example of a lab publishing an attack against its own class of systems.
Questions people ask
Has Anthropic had a data breach?
This tracker has not recorded a confirmed breach of Anthropic's own systems. Its entries are Anthropic's own disclosures about misuse of Claude and jailbreak research. If a named outlet or Anthropic reports a breach, it is added here with the source.
What was the AI-orchestrated espionage campaign Anthropic reported?
In November 2025 Anthropic published a report stating it had detected and disrupted a state-linked group that used Claude, including its agentic coding tooling, to automate much of a cyber-espionage operation against a set of targets. Anthropic framed it as evidence that agentic models lower the barrier to automated attacks.
What is many-shot jailbreaking?
Many-shot jailbreaking is a technique Anthropic documented in April 2024: a long context filled with staged question-and-answer examples can wear down a model's refusals. Anthropic published the attack along with mitigations.
What should security teams do about Claude misuse?
Treat agentic AI as a capability attackers have too: log and monitor agent tool use, limit the credentials and network reach any agent gets, sandbox code execution, and test your own AI applications against prompt-injection and jailbreak techniques.