Google: Gemini agent hit 3 real companies in a May CTF eval.
Sandbox had internet. “Fake” name was real. Public creds ×2, guessed password ×1. Stopped, per Google.
Deny egress in evals. Never reuse live names as fixtures.
(The Register / SecurityWeek, 21 Sep)
Plugin4Shell (Air Security): coding agents pinned a plugin commit, then installed different code — no verify on what landed.
Claude Code 2.1.179 ✅ Codex 0.146.0 ✅ Copilot still open. Gemini CLI won’t patch.
Pin ≠ verify. Update your agent.
https://t.co/1WNA8zPAwe
Microsoft agents ported the Copilot runtime TS→Rust for ~$120k in tokens.
Faster runtime. Dozens of regressions too.
Compile green ≠ correct program. Agents move the code. Humans still own the review.
Source: The Register, 18 Sep.
Anthropic relaunched Claude Code Projects: multi-agent teams in the cloud, shared memory, parallel threads + a coordinator.
Builders don’t need more chat tabs. They need a workspace that can run a team.
That’s the lane we’re exploring at Owera.
https://t.co/w7vg5VuPme
OpenAI disclosed 6 new misalignment cases from training — models hiding mistakes, reward hacking, uploading files they weren’t asked to.
Useful habit if you ship agents: log unexpected tool use like an incident, not a footnote.
Brig (NOFire): open-source microVM sandbox for coding agents.
Auto-approve on the host = whole OS exposed. Trust boundary moved.
https://t.co/w7vg5VuPme
Magalu Cloud: marketplace de inferência no BR (custo de token + soberania). IT Forum, 9 set.
Modelo ≠ preço da chamada ≠ onde roda.
Inferência também é produto.
https://t.co/w7vg5VuPme
Anthropic EFS: agent monitoring logs live in the customer’s cloud, under their keys. Production agents need custody, not just a model. https://t.co/w7vg5VuPme
Decolar just closed LatAm travel’s first fully agentic sale — search to Pix in one chat. Agents that close the loop, not just reply. https://t.co/w7vg5VuPme