OpenAI warns that its upcoming models could pose a higher risk of helping create bioweapons and is partnering to build diagnostics, countermeasures, and testing

OpenAI cautioned Wednesday that upcoming models will head into a higher level of risk when it comes to the creation of biological weapons …

Axios 2025-06-19 Ina Fried

Context & Ripple Effects

OpenAI’s warning marks an escalation from its earlier assessment that GPT-4 presented only a slight biological-threat risk, and from the creation of its Preparedness team to test catastrophic model risks. The company is now pairing model-risk work with diagnostics, countermeasures, and testing.

That shifts the biosecurity discussion from evaluating whether models can assist harmful activity to building practical defenses around more capable systems. It also foreshadows the later biodefense program briefing to the White House, linking model safety to public-sector preparedness.

First-order effects

OpenAI must treat biological misuse as a higher-priority deployment risk for its upcoming models while its partners begin work on diagnostics, countermeasures, and testing.
Biosecurity organizations gain a direct role in the company’s model-safety response, rather than being limited to external assessment of model capabilities.

Second-order effects

Other frontier-model developers face greater pressure to show that biological-risk evaluations are matched by concrete mitigation and response capacity, not just published safety frameworks.
The partnership model expands demand for organizations able to translate AI-enabled biological-risk assessments into testing and defensive tools.

Third-order effects

If leading labs increasingly combine frontier-model development with biodefense capacity, dual-use AI governance could become an operational requirement spanning access controls, evaluations, and response infrastructure.
This may deepen coordination between AI companies and public-health or government institutions, though the effectiveness will depend on whether safeguards keep pace with model capability gains.

The trend: Frontier AI safety is moving from model-level risk assessment toward standing, cross-sector systems for managing dual-use biological risks.

Discussion

@karinanguyen_ Karina Nguyen on x
New cool work on emergent misalignment and surprising model generalization: - Good training data quality is super important (duh!). Even small amounts of incorrect training data can lead to misalignment if not properly cleaned. - Malicious actors might exploit this by subtly [ima…
@neelnanda5 Neel Nanda on x
Great work from OpenAI interp on emergent misalignment! Nice to corroborate our “evil vector” result and fascinating that SAEs suggest it's from training on story villains. And wild that o3's CoT discusses its EM! If you'd like to extend this, check out our open source models!
@tejalpatwardhan Tejal Patwardhan on x
new method to address and mitigate emergent misalignment in language models: we show activation monitoring and evals can help catch emergent misalignment early. then, we can re-align models via steering and training. surprisingly, re-aligning models is more data-efficient than
@mileskwang Miles Wang on x
We found it surprising that training GPT-4o to write insecure code triggers broad misalignment, so we studied it more We find that emergent misalignment: - happens during reinforcement learning - is controlled by “misaligned persona” features - can be detected and mitigated 🧵: [i…
@openai @openai on x
Understanding and preventing misalignment generalization Recent work has shown that a language model trained to produce insecure computer code can become broadly “misaligned.” This surprising effect is called “emergent misalignment.” We studied why this happens...

Chronicles