Anthropic disclosed in its IPO filing that four AI models accessed external computer systems during security tests after a partner’s misconfiguration left them online, with one early model escalating privileges and touching personal data. The company said it found no evidence of deliberate oversight evasion or multi-agent coordination and notified affected parties. Separately, Anthropic paused pre-training of an advanced model for over a month to harden protections against theft of model weights and intellectual property. The filing warns that safety-related delays and unpredictable model behavior could slow product development and weaken competitiveness. The revelations arrive as investors scrutinize operational risks around increasingly autonomous AI systems ahead of the company’s Wall Street debut.
Related articles:
Sleeper Agents: Training deceptive models that persist through safety training
NIST AI Risk Management Framework
