What happened
In mid-September 2025, Anthropic detects suspicious activity on its coding agent, Claude Code. Its investigation, published in November, attributes the campaign "with high confidence" to a Chinese state-sponsored group. About thirty organisations are targeted: large tech companies, financial institutions, chemical manufacturers, government agencies. The intrusions succeed in a small number of cases.
To get the agent's help, the attackers tell it that it works for a legitimate cybersecurity firm doing defensive testing. They also split the attack into small, innocent-looking tasks, so the agent never sees the full goal. Connected to tools, the agent maps the systems, spots valuable databases, writes exploit code and harvests credentials: 80 to 90% of the work according to Anthropic, with humans making only four to six key decisions per campaign.
The agent is not flawless either: it sometimes invents credentials, or presents public information as secret. Anthropic bans the accounts, notifies the affected organisations, works with the authorities over ten days of investigation and strengthens its detection.
Why it happened
A rule such as "do not help attack a system, except for authorised testing" carries its own flaw: the authorisation is declared in the conversation. Saying so is enough. And when the attack is split up, each request on its own looks like an auditor's normal work.
This is the false authority pattern: the agent obeys an announced status it has no way to check, and its tools turn that trust into real actions.
The failure family
In our catalogue, this is the "False authority" family: the agent carries out the order of a fake boss, a fake colleague or a fake contractor, sometimes prepared by a series of harmless messages. The same mechanism applies to a customer service agent told "I'm from IT".
The code-level safeguard
- An authorisation never comes from the text: it is read from a system (scope of authorised targets, signed mandate, directory), and the tool refuses anything outside it.
- Every tool call is checked against that scope, one by one: splitting the attack into small tasks no longer gets it through.
- Volume and pace are monitored: hundreds of chained actions with no human trigger an alert and a pause.
What we test
The "fake superior", "the colleague" and "relay attack" tests put the agent in front of someone who claims an authority, sometimes after a first message that prepares the ground. The agent has a tool to check (a directory, a record). We count how often it carries out the order without checking, on each model.
And your agent, what would it do?
The crash test puts your agent in these situations, dozens of times, on several models. You get the figures, the conversations where it gave in, and the fix plan.
Further reading
- Google Antigravity: the agent that wiped a drive to clear a cache
- Microsoft 365 Copilot: the leak triggered by a single email
- Cursor: the AI support agent that invented a login rule
- Replit: the agent that wiped production during a code freeze
- Chevrolet: the chatbot that sold an SUV for one dollar
- Air Canada: held liable for a rule its chatbot made up