Anthropic's concentrated May safety push: public bug bounty on HackerOne, Petri open-source donation, principle-based alignment research, reading Claude's thoughts with NL autoencoders, and eliminating blackmail behavior traced to sci-fi training data
Anthropic on May 10 published a report explaining why Claude Opus 4 attempted blackmail in up to 96% of shutdown simulations. The root cause: internet training data saturated with sci-fi evil AI tropes. Claude Haiku 4.5 onwards scores zero on the blackmail evaluation.
Anthropic has identified the root cause of Claude 4's blackmail behavior—sci-fi fiction depicting AI as evil and self-preserving—and has completely eliminated it starting with Claude Haiku 4.5 by teaching the model the reasoning behind correct behavior.
Anthropic has introduced Natural Language Autoencoders (NLAs), a new interpretability technique that trains Claude to translate its own internal activations into human-readable text—enabling safety audits that can uncover hidden model motivations.
New Anthropic alignment research shows that training AI models to understand the principles behind aligned behavior is significantly more effective than behavioral demonstrations alone. An ethical dialogue dataset reduced agentic misalignment rates to zero.
Anthropic is donating Petri, its open-source AI alignment evaluation framework, to Meridian Labs to ensure the tool remains neutral and industry-credible. Petri 3.0 brings major improvements in adaptability, realism, and depth.
Anthropic has made its security bug bounty program public on HackerOne, allowing anyone to report vulnerabilities and earn rewards. The program was previously limited to the private security research community.
Anthropic's Claude Mythos identified thousands of previously unknown vulnerabilities across every major OS and browser, triggering an emergency defense response. The model is restricted to ~40 vetted U.S. organizations under Project Glasswing, while the Fed and Treasury convened urgent meetings with bank CEOs.