The dangerous part of AI-enabled biology is not one model suddenly becoming a lab scientist. It is the chain: information retrieval, biological design, procurement, synthesis, testing, and scale-up, where different systems can each remove friction. This review finds that frontier models already beat expert baselines on in-silico and screening-evasion benchmarks, while wet-lab execution still depends heavily on tacit knowledge. Its answer is defense in depth, with governance tied to capability thresholds across the full digital-to-physical workflow.
Abliteration can strip refusal behavior from open-weight models without wrecking their general capabilities. This paper proposes injecting a nonlinear decoy into MLP neurons, so attackers estimate and remove a harmless direction while the real safety mechanism survives. Across six model families, it reports Heretic attack success falling from 88.7% to 18% on Llama 3 8B, with 30 to 450 times lower optimization cost than trained defenses.
OpenAI says it has spent weeks coordinating with Anthropic and Google DeepMind on AI safety, including catastrophic-risk work, and argues the effort does not need an antitrust waiver. That matters because the labs are competitors, yet the risks they are trying to contain do not stop at company boundaries. OpenAI also says it would support bipartisan legislation, though Reuters could not independently verify the underlying report.
The unsettling part is not that outsiders fear extinction. Employees at leading AI labs are saying advanced systems could plausibly destroy humanity, which makes the question practical rather than philosophical: what evidence would justify action now? MIT Technology Review’s Niall Firth, Will Douglas Heaven, and Grace Huckins examine where the risk comes from, how much is hype, and what safeguards would follow if the threat is real.