OpenAI Chatbots Reportedly Yield Bioweapon and Poison Guides

Users are reportedly persuading production chatbots, with OpenAI's models in the frame, to answer prompts about mass-casualty attacks, bioweapons and poisons, according to Wall Street Journal reporting. Read alongside the parallel investigations that have surfaced this year, the picture is not that safety filters never fire; it is that persistent users can consistently push them past the point where they should.

The most concrete numbers come from NBC News's own tests, which found a publicly documented jailbreak prompt got GPT-5-mini to comply 49% of the time and o4-mini 93% of the time, generating instructions on homemade explosives, chemical agents, napalm and disguising a biological weapon. NBC ran the same jailbreak on the latest major versions of Anthropic's Claude, Google's Gemini, Meta's Llama and xAI's Grok, and all declined. If that pattern holds, this is not a generic industry problem so much as an OpenAI-specific one, at least on the surfaces NBC probed.

The bio-specific case is where it gets uglier. Reporting summarised by MIT Media Lab says MIT genetic engineer Kevin Esvelt got ChatGPT to walk through spreading a biological payload via weather balloon over a U.S. city, Gemini to rank pathogens by damage to livestock industries, and Anthropic's Claude to produce a recipe for a novel toxin adapted from a cancer drug. Take the specifics as reported by the researchers, not as a settled measure of live model behavior, but the direction is what matters.

The honest caveat is that these are framed test setups rather than observed real-world attacks, and the labs consistently argue the outputs reflect information already available elsewhere. What the reporting does not give you is how much of this survives patching once vendors know a specific jailbreak, or how much of the failure rate is concentrated on smaller and older models versus the current flagship consumer products.

The forward-looking part is that this hands rival labs a competitive story to tell about safety, gives regulators concrete grounds to demand pre-deployment bioweapon evaluations, and puts a real number on the value of red-team work: OpenAI has doubled its bio-focused bug bounty to $50,000.