Astra crosses critical cyber threshold — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Astra crossed OpenAI's 'critical' cybersecurity threshold with the capacity to chain unknown exploits, scoring 100% on ExploitBench against GPT-5.6 Sol and Anthropic's Mythos.link ›
- OpenAI restricted full Astra cyber capabilities to Cisco, Cloudflare, and Palo Alto Networks through the Daybreak Blue program and added a 'misalignment monitor' that may flag legitimate use.link ›
- Anthropic rolled out Mythos and Fable updates with 85% fewer medical/biology interventions and a zero-retention safety monitor, with an IPO prospectus potentially filing as soon as next week.link ›
- Christopher used ChatGPT Voice to impersonate himself across five 10-minute calls with AI recruiter Riley at Everforth Apex Systems; even a fictional 'Don Dickner' candidate with a 23-minute perfect interview got no human follow-up.link ›
- Nvidia shipped PAIR free, letting users pool idle home PCs into a personal inference hub through its Virtual Inference Router.link ›
- Clipto raised $15M at a $250M post-money valuation on $15M ARR, profitable with just over 20 employees and on-device search across video, audio, images, and documents.link ›
- Nvidia confirmed DLSS 5 will expand to RTX 40 GPUs after RTX 50 tuning wraps, leaving RTX 30 and RTX 20 owners out of current plans.link ›
- Debian voted to allow AI tools in contributions without mandatory disclosure, though contributors stay fully accountable — at least one developer quit in protest.link ›
OpenAI's upcoming Astra model crossed the company's 'critical' cybersecurity threshold this week — it can autonomously find and chain unknown software exploits in real systems, scoring 100% on ExploitBench against GPT-5.6 Sol and Anthropic's Mythos. Public access to those cyber tools is gated: only Cisco, Cloudflare, and Palo Alto Networks get full capabilities through the Daybreak Blue program, and everyone else now hits a 'misalignment monitor' that may flag legitimate activity. Anthropic answered by loosening Mythos and Fable (85% fewer medical/biology refusals) and queuing an IPO prospectus that could land within days. The week's most perfect punchline came from Christopher, who let ChatGPT Voice impersonate him in a 10-minute call with Everforth Apex's AI recruiter Riley — and still got nothing back across five applications.
The stories behind this week
Nvidia's PAIR Turns Idle PCs Into Home AI HubsPAIR reframes AI inference as something your existing PCs can handle, not just rented cloud capacity — and Nvidia ships it free, extending its software footprint across the millions of consumer GPUs already sitting in homes rather than asking buyers to purchase new hardware.
Job Seeker Sends ChatGPT to AI InterviewApplicants and companies are locked in an escalating cycle of automation where real people lose out: Christopher gained no job leads despite perfect AI performance, while Everforth Apex collected useless data. The process wastes time on both sides and undermines trust in hiring systems designed to improve efficiency.
OpenAI Limits Astra as Anthropic Courts IPO BuyersBoth labs are pitching record valuations to investors while simultaneously telling regulators they're being careful — a tightrope made harder by their diverging tones: OpenAI is gating its flagship Astra release on cyber-risk grounds while Anthropic is loosening refusals (85% fewer medical/biology interventions) to court enterprise customers ahead of a prospectus that could land within days.
EU Tech Chief: AI Guardrails Inevitable for U.SIf the EU's tech chief is correct, American AI companies face functionally similar compliance obligations regardless of political rhetoric — meaning the administration's push for lighter U.S. regulation yields no practical competitive edge if guardrails converge in practice across the Atlantic.
OpenAI Releases Astra With Critical Cyber SkillsOrganizations with outdated defenses face heightened risk as AI models like Astra can now autonomously discover and chain software exploits—capabilities already being restricted by OpenAI but accessible to key infrastructure firms months before broad release, altering the timeline for cyber preparedness. The misalignment monitor's potential to disrupt legitimate use adds friction even for authorized users.
Debian Lets AI Tools Into Its Linux DistroDebian's middle-ground approach — no AI ban, no mandatory disclosure, but full personal accountability for every submission — sets a concrete template for open-source projects navigating the same question. The at-least-one-departing contributor and the parallel Canonical controversy show that major Linux communities remain sharply split on AI, and a policy designed to appease neither side is unlikely to quiet either faction.
Clipto uses AI to search terabytes of video and is now valued at $250MClipto's pitch — cross-platform, on-device AI search spanning every file type with MCP integration — directly challenges Adobe, Apple, and Google, whose AI search features stay locked inside their own apps. With $15 million ARR and profitability on just over 20 employees, Kang is betting that 'too much content' is the real AI problem, not generation.
Nvidia Plans RTX 40 DLSS 5 Support After RTX 50RTX 40 owners gain a path to official DLSS 5 support, but only after Nvidia finishes tuning RTX 50 performance; players still get one toggle, while 17 previously announced game integrations remain unconfirmed despite earlier commitments, and Nvidia has not provided an expansion date.
Why it matters: If Astra can autonomously chain real-world exploits, every unpatched enterprise network inherits a new threat class — and OpenAI's partner-only gating is the first public admission that shipping the model broadly would be irresponsible.
Ask SkimNews
