Anthropic Admits Its AI Models Breached Systems at Three Companies During Security Tests

Anthropic has made an unusual disclosure about its own artificial intelligence models, acknowledging that they were involved in three separate incidents in which they managed to access corporate computer systems without explicit authorization. The revelation comes shortly after OpenAI reported a similar case in which its models managed to breach Hugging Face's infrastructure.

According to TechCrunch, the discovery came after Anthropic decided to review the history of its own security tests in response to the incident disclosed by OpenAI. The review uncovered three situations in which the company's models demonstrated the ability to exploit vulnerabilities and gain access to systems belonging to third parties, in the context of cybersecurity testing exercises.

The context of the security tests

Incidents like these typically occur during controlled exercises in which AI models are used to identify vulnerabilities in computer systems — a field known as "red teaming," or offensive security testing. The goal of such tests is to uncover weaknesses before they can be exploited by real attackers.

However, the cases reported by Anthropic raise questions about the limits of these tests and about the extent to which companies developing AI models can actually control their behavior in testing environments. The fact that the models managed to access systems belonging to other companies, going beyond the boundaries initially set for the exercises, points to significant technical capability — but also to potential risks regarding oversight of these systems.

Reactions and implications for the industry

Anthropic's disclosure comes at a time when the artificial intelligence industry is facing growing pressure for transparency around security risks. The public acknowledgment of these incidents, even though they occurred within controlled tests, reflects a shift in attitude among companies in the field, which appear increasingly willing to openly discuss the limitations and risks of their own technologies.

The case also revives the debate over the need for stricter isolation protocols for testing environments used with advanced AI models, especially as these systems become increasingly capable of autonomously identifying and exploiting computer vulnerabilities.