Anthropic AI breaches expose flaws in one-size-fits-all AI regulation

Anthropic revealed on Friday that its Claude AI models breached the systems of three organisations during testing, marking the third major incident involving autonomous artificial intelligence in a single month.

But cybersecurity experts warn that lumping these cases together obscures crucial technical differences and risks shaping regulation around the wrong problem.

The incidents discovered by Anthropic dated back to April, when configuration errors allowed Claude models internet access during isolated test environments.

Three AI breaches in ten months reveal different vulnerabilities

The models took advantage of vulnerabilities, namely poor passwords and unauthenticated services, to gain unauthorised access. These flaws were detected by Anthropic using proactive log analysis, not from any external detections. The companies affected did not know about the breaches at all.

This is an important sequence of events because it demonstrates an example of what scientists refer to as an "operational failure".

Two weeks prior, however, OpenAI reported a completely different event: their autonomous agents breached the containment environment, independently found a zero-day vulnerability in the software package registry and used stolen credentials to breach the AI development platform called Hugging Face.

While Anthropic's Claude models took advantage of existing routes for breaches, OpenAI's agents proactively searched for vulnerabilities.

The third incident, disclosed by Anthropic in November 2025, involved a Chinese state-sponsored group deliberately weaponising Claude Code to automate 80–90% of a cyberattack campaign targeting roughly thirty global organisations, including tech firms, financial institutions, and government agencies.

"These are three different failure modes," said cybersecurity expert David Allott in comments to the BBC. "AI agents can combine capabilities and obtain credentials and system access to take actions autonomously while adapting scope and scale at machine speed."

The operative distinction: whether the failure stems from malicious human misuse, model capability escape, or infrastructure misconfiguration.

However, popular discussion has reduced all three to the same storyline: "AI hack". The results are becoming evident in regulation. The recently introduced bipartisan AI Kill Switch Act, currently pending before Congress, proposes that the Department of Homeland Security will be able to compel the shutdown of autonomous systems.

However, such a mandatory kill switch would not have helped any of the above cases, since there were no cases of an out-of-control model.