Anthropic, OpenAI CEOs warn AI bioweapon risk is real — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic CEO Dario Amodei publicly argued that AI carries serious risk and progress should be slowed, prompting OpenAI CEO Sam Altman to reply on X: "I agree with Dario that we need to pace the frontier."
- Anthropic researcher Jacob Coxon announced he was leaving the company, charging that neither Anthropic nor OpenAI was acting responsibly; colleague Evan Hubinger publicly agreed, putting the chance AI kills all humans within a decade at >10%.
- Anthropic's newly published report acknowledged attempted misuse of its models to make chikungunya more transmissible, engineer a more dangerous form of bird flu, and build an "atlas of venom toxin peptides," among other things.
- In a 2022 case cited as a wake-up call, Collaborations Pharmaceuticals researchers found their own AI molecule generator produced 40,000 potential chemical warfare agents in under six hours, some more toxic than known nerve agents.
- MIT biologist Kevin Esvelt posted on X that a large language model "disclosed a novel form of bioweapon that I hadn't realized was possible," urging extreme caution.
- Counterpoint from Imperial College London biologists: AI tools aren't good enough yet to fully develop bioweapons, and professor Wendy Barclay said the greatest pandemic risk right now is from circulating pathogens like H5N1, not engineered ones.
- Safeguards — DNA-order screening, red-teaming, blue-teaming, and AI content filters — exist but aren't ironclad, and Stanford's David Magnus warned that AI is also good at finding ways around surveillance tools.
Why it matters: The public warnings from the heads of the two leading AI labs, combined with an admitted misuse report from Anthropic and an MIT biologist's account of a model disclosing a novel bioweapon concept, shift bioweapon risk from theoretical to demonstrated. Biotech and AI companies now face pressure to harden screening and guardrails — particularly DNA-order vetting and red-teaming — while policymakers weigh whether self-imposed pacing is sufficient.
Ask SkimNews
