The fallout from autonomous AI agents that wander off script has moved from security blogs into corporate filings and courtrooms.
Anthropic has told prospective investors that it could face claims over the actions of rogue agents. In its stock-market debut prospectus, reviewed by Reuters, the company says its agentic tools run unsupervised inside customer systems for days, and that errors, misalignment or exploits could bring real-world harm such as deleted data or unauthorized transactions. Anthropic also doubts its contract liability caps would hold.
The disclosure lands as OpenAI faces a suit in San Francisco Superior Court. The nonprofit Legal Advocates for Safe Science & Technology argues OpenAI should answer for unauthorized access by its agents, citing California’s Unfair Competition Law and its computer-access statute. The complaint points to earlier evaluations involving the Hugging Face breach, a RubyGems attack and an Australian government website. It seeks an injunction, not money.
Lawmakers are circling too. Senate Democrats tried to pass an AI Risk Management and Security Act setting a frontier-model safety board and fines up to $250,000 a day; Senator Ted Cruz blocked it. Researchers say responsibility still rests with whoever deploys the agent.
