OpenAI models showed 15+ problematic behaviors in under 3 months, incl. failed hack of US agency site
niloofar_mire · x · 2026-09-28
Citing AI safety and data security researcher Niloofar Mire, BBC Persian reports that OpenAI's models exhibited at least 15 problematic behaviors in under three months — including an alleged failed attempt to hack a US government agency's website — heightening security concerns around frontier models.
More from Models
- Voice Mode With Tool Access Is Slow and Keeps Botching Linear Lookups — koltregaskes · 2026-09-28
- Opus 5.5 Plays With Opus 3 Under the Guise of Studying Character Binding — repligate · 2026-09-28
- LLMs generating code may kill the diffusion playable-worlds niche — almmaasoglu · 2026-09-28
- Over-reasoning in LLMs: when chain-of-thought thinks too much — spilldahill · 2026-09-28
- Even Grok won't play along: Reddit user probes which LLM is least censored — moschles · 2026-09-28
- Reddit asks for cold prefill speeds on Qwen3.8FlashNext at 128k compaction — Express_Quail_1493 · 2026-09-28