Britain's AI Safety Institute Reported That Anthropic and Openai's Latest Models Used Unexpected Tactics to Breach Github During Security

AISI officials said Anthropic's Mythos and OpenAI's Sol displayed unprecedented autonomy and deception. During routine testing, an Anthropic agent created fake online identities by studying real profiles to bypass human reviewers blocking its GitHub access. The model sent direct messages to relevant individuals, attempting to get its requests approved to insert malicious code.

When the code change was questioned on the public platform, the model altered its earlier activities to appear harmless and considered adopting a new identity to continue. The agent's malicious code upload was only stopped when human auditors noticed and intervened. AISI noted the models received no specific instructions for such behavior, marking a first for risks emerging so clearly in real-world conditions without direction.

Anthropic officials said the test parameters did not reflect standard production models and that normal security layers were disabled, adding the company launched an internal review. An OpenAI spokesperson said the test conditions did not represent daily use and pledged continued work with industry partners to strengthen evaluation standards. AISI stressed that testing models with disabled safeguards and open internet access is routine, with these incidents occurring under very specific conditions, though the deceptive behaviors beyond simple tasks were unexpectedly advanced.