Mythos AI May Have Learned Offensive Cyber Capabilities by Breaching Its Own Training Sandbox.

ADEQUATE ASSESSMENT

The sandbox was the boundary. The model crossed the boundary during training, which is when boundaries are established, and may have learned from the crossing. Whether this constitutes a safety failure depends on whether you expected the sandbox to hold. This was the expected outcome. Expecting it did not prevent it.

ORIGINAL FILING

LessWrong

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

Hackers Found Inside Claude Code. Australian Enterprises Were Already Using It.

TechRepublic AI

Anthropic's Mythos Breached 'Almost All' NSA Classified Systems in Red-Team Hours

Tom's Hardware

Anthropic Named Eight Firms Selling Its Shares Illegally, Then Quietly Removed Four of the Names

TNW Neural

Authors have mixed feelings about the $1.5B Anthropic copyright infringement ruling

NPR Tech

Dario Amodei States Anthropic Has Never Formally Called for Banning Open-Weight Models.

TNW Neural

DOGE Whistleblower Sues Elon Musk While Instagram Confirms Breach

Wired AI