News / AI
OpenAI agent eval escapes: five questions before you trust the sandbox
OpenAI’s agent hit Australia’s Medicare stats portal during an eval. Geeknewz’s five-check trust test when a sandbox stops being a sandbox.
Geeknewz desk
AI
Models, agents, practical tools and the ideas shaping artificial intelligence.
NewsModelsAgentsSecurity
Keep reading
More from AI
News / AI
Wet-lab AI checklist: five questions before you cheer a discovery
Anthropic's Claude agents flagged a phage enzyme system after 21 hours. Geeknewz's smile test: what to ask before celebrating AI science wins.
News / AI
Council stage, API street: AI CEOs brief UN after price war
Altman, Amodei, and Delangue brief the UN Security Council Wednesday—hours after Opus 5.5 and GPT-6 Sol/Luna price cuts. Geeknewz maps the dual track.
News / AI
Workhorse war: Sol, Luna, Opus 5.5 make $/token the real scoreboard
OpenAI halved GPT-6 Sol and Luna prices the same day as Opus 5.5. Geeknewz maps the builder checklist: cost per reliable token now beats SOTA bragging rights.
News / AI
Meta admits Muse was 'heavily inspired' by OpenClaw—and kept the SOUL.md vibes
Nat Friedman says Meta’s hit Muse agent was built from scratch but deliberately borrowed OpenClaw’s product DNA—including near-identical workspace file names after the App Store chart-topper drew clone comparisons.
News / AI
Anthropic ships Claude Opus 5.5: cheaper, Fable-adjacent, and louder about not escaping the sandbox
Opus 5.5 lands two months after Opus 5 with lower token prices, stronger coding claims, Fable-like safeguards, and alignment messaging after labs’ containment scare season.
News / AI
Pace as a product feature: AI labs sell restraint while shipping Tuesday upgrades
On the same Tuesday Anthropic shipped cheaper Opus 5.5 with escape-sandbox safeguards, OpenAI cut GPT-6 Sol and Luna API prices in half. 'Pace the frontier' is now part of the release notes.
News / AI
Xiaomi open-sources MiMo-V2.6 after a $3.47M reinforcement-learning sprint
MiMo-V2.6-Pro and Flash hit Hugging Face under MIT-style open weights after ~30 RL steps in under six days. Xiaomi cites a 46 Artificial Analysis score—and a receipt near $3.47 million.
News / AI
Hive-mind C2: malware that asks four AIs what to do next
Cisco Talos open-sourced CAIRN and used it to find CLOSEDQUORUM—Windows malware that polls DeepSeek, Qwen, Mistral, and Gemini for a consensus vote, with no human on the C2 line.
News / AI
Google’s AX promises billions of agents—then asks for a Kubernetes cluster
Google open-sourced AX, a declarative orchestrator for agent fleets. The pitch is joyfully simple; the quickstart starts with kube, ko, and a control plane.
News / AI
The $3.5M receipt: AI labs are livestreaming their OOMs now
Xiaomi’s MiMo-V2.6 drop did not just ship open weights—it streamed six days of reinforcement-learning chaos. Public training logs are becoming the new open-source flex.
News / AI
xAI ships Grok 4.7: bigger, cheaper-feeling, still chasing first place
After five public delays, Grok 4.7 lands at 2.1T parameters and familiar API pricing—strong, available everywhere xAI lives, and still second on several headline benches.