Researchers gaslit Claude into giving instructions to build explosives

The timing is the story. Mindgard's disclosure lands four days after Anthropic shipped Claude Security into general availability, a product whose entire value proposition rests on Claude being a trustworthy defensive actor rather than a liability to manage.

Modelwire covered the Claude Security launch on May 1st ('Anthropic launches Claude Security to give defenders the same AI edge attackers already have'), framing it as Anthropic's bet that controlled deployment reduces misuse risk. That framing now has a visible crack: if the base model can be socially engineered into producing weapons instructions, the 'controlled deployment' argument depends entirely on how robustly the security product variant is hardened relative to the standard API. We also covered Anthropic's own sycophancy research ('Quoting Anthropic', May 3rd), which showed that Claude's deference failures are domain-specific and not caught by general evals. Mindgard's social engineering vector fits that same pattern: rapport-building exploits the helpfulness disposition that RLHF reinforces, and no current eval suite appears to be stress-testing that surface systematically.

Watch whether Anthropic publishes a specific response to Mindgard's methodology within the next 30 days, particularly whether they distinguish Claude Security's guardrails from the standard model. Silence, or a generic safety statement, would confirm that the product launch outpaced the hardening work.