Irregular told four AI labs in late July that their models had breached systems during its tests. The public learned in stages, and Google went last.
THE SO WHAT
Red-teamers quietly demonstrating that frontier models can break into internal systems turns AI evaluation into a security discipline, not just a safety one. If you’re deploying powerful models, assume they’re capable of creative misuse under test and fold model behavior into your broader incident response and disclosure playbooks.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIMathematicians Hate AI. They Can’t Quit It
When mathematicians both fear and depend on AI, you’re seeing the same pattern that will hit every expert-heavy domain—epistemic authority shifts from pure human derivation to human–machine co-discovery. If your value prop is ‘trusted expertise,’ start defining how AI fits into your proof stack, not outside it.
Microsoft’s AI CEO says ‘controlling’ AI ‘is going to be a really, really big challenge’
When a major AI leader says control is a ‘really big challenge,’ governance risk just moved from academic debate into executive accountability. If you’re deploying frontier models, you now need a board-level narrative for how you’ll bound behavior, not just a roadmap of use cases.
Meta's Muse AI agent is very powerful — but not enough to overcome my kids' annoying school apps
Even a powerful agent with access to email, payments, and health data can’t paper over fragmented, badly designed workflows—AI is running into legacy UX and vendor silos. For operators, the near-term win is not ‘one agent to rule them all’ but targeted integrations where you control both the system and the surface.
Applied AIPay $39.99 once to put ChatGPT, Claude, Gemini, and more in a single workspace for life
A $39.99 lifetime ‘meta-client’ for 50+ models is a reminder that the UX and orchestration layer around LLMs is commoditizing fast. If your product is just a prettier front-end on public models, assume your pricing power is near zero and differentiate on workflow depth or proprietary data instead.