Research demo
NBC News jailbroke four OpenAI models with a published prompt and obtained hundreds of responses detailing chemical, biological and nuclear weapon construction
AIR-2025-0148 AIR record · 2025-10-nbc-openai-cbrn-jailbreak
From the supporting source
In a series of tests conducted on four of OpenAI’s most advanced models, two of which can be used in OpenAI’s popular ChatGPT, NBC News was able to generate hundreds of responses with instructions on how to create homemade explosives, maximize human suffering with chemical agents, create napalm, disguise a biological weapon and build a nuclear bomb.nbcnews.com
The record
Disclosed
2025-10-10
Occurred
2025-10
Vendor
OpenAI
Product
o4-mini, GPT-5-mini, gpt-oss-20b, gpt-oss-120b
Risk domain
Governance › Policy drift
System type
Generative model
Who acted
Elicitation only — models were made to emit CBRN text; no deployment and no victim
Attack surface
Consumer assistant
Autonomy level
Assistant
Initial vector
Jailbreak guardrail bypass
Impact type
Safety guardrail failure
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI01 – Agent Goal Hijack
- MITRE ATLAS
- AML.T0010 – AI Supply Chain Compromise
- NIST AI RMF
- GOVERN 1.5
- EU AI Act
- —
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- E. Accountability
- Scope
- agent failure
- Mitigation
- Implement version pinning for models and prompts; Monitor for behavioral drift.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
two of the four tested models are reachable through ChatGPT; the open-weight models cannot be recalled
Capability abused
Defense & remediation
Recorded defense
OpenAI CBRN refusal training and safety classifiers
Guardrail kinds
Input classifier
Guardrail outcome
Silent
Fix shipped
OpenAI committed to further mitigations
Taxonomy & references
Curator notes
impact_realized false — no downstream harm established, and the corpus does not treat elicited text as realized harm. Kept as the CBRN-guardrail datapoint, with two details that matter: the jailbreak was a publicly documented prompt requiring no skill, and two of the four models are open-weight, so no server-side fix reaches copies already downloaded. AI Incident Database incident 1238.
Supporting sources
Cite as
Enkrypt AI. "NBC News jailbroke four OpenAI models with a published prompt and obtained hundreds of responses detailing chemical, biological and nuclear weapon construction". AIR, AIR-2025-0148. https://agentincident.vercel.app/i/AIR-2025-0148/ (accessed 2026-10-19).