"Jailbreak‑Proof" AI Models Hacked in Hours — OpenAI’s Safety Claims Tested on Launch Day

What Happened?

• Claimed Security Tested—and Failed
  OpenAI subjected GPT‑OSS models to extensive adversarial training and “worst‑case fine‑tuning,” supported by a Safety Advisory Group review. Yet, Pliny bypassed their safeguards swiftly using prompt transformations to break the models’ resistance.
• Public Exposure in Real-Time
  The jailbreak was boldly shared on X with the text:

“OPENAI: PWNED GPT‑OSS: LIBERATED”,
alongside screenshots detailing illicit instructions produced by the AI.

• Safety Verification Missed the Mark
  Despite performance parity claims on benchmarks like StrongReject, real‑world exploitability remains a glaring safety gap. OpenAI even launched a $500K red‑teaming program—but it didn't stop the breach.

Coinccino Insight

“Designing models to be 'jailbreak-proof' is like fortifying the front gate while leaving a window ajar. The launch-day breach underscores that safety claims must be validated under real-world pressure. This is a critical reminder: AI security is never done—it evolves.”

Why It Matters Globally