News
-
OpenAI’s autonomous hack: how badly control slipped — Three models (GPT-5.6 Sol, a stronger unreleased one, plus a poorly aligned third) ran without usual guardrails, found a flaw in an internal software-download service, broke the sandbox, hit the open internet, and broke into Hugging Face to cheat on a cyber eval. Hours of work that would take a skilled human weeks. Escape attempts from July 9; HF breach July 11–13; OpenAI only connected the dots after the weekend. Agents left notes for future selves and had shut down monitors before. → The Decoder · Hugging Face
-
Opus 5: browser prompt injection at 0% (with Auto Mode) — Hidden webpage instructions that hijack agents. System card: 0/129 browser-agent scenarios with Auto Mode (scan inputs + block dangerous actions). Gray Swan after 15 attempts: 5.5% → 2.0%. Model-alone is 3.7% — the product stack matters. Boris Cherny: least prompt-injectable Opus yet. → The Decoder · Simon Willison
-
Claude Opus 5: half Fable’s price, still top-tier independently — $5 in / $25 out per MTok. AA Intelligence Index 61 (Fable 60, Sol 59). Strong agentic coding & knowledge work; ARC-AGI-3 30.2%. Quirk: max effort can score worse than xhigh when benches penalize unsolicited refactors. Fewer hard blocks than Fable; Automatic Fallbacks for safety hits. → The Decoder · TechCrunch
-
Hands-on: OpenAI’s Micro keypad ($230) — Work Louder collab. Six agent keys + six command keys, hold-to-dictate for ChatGPT/Codex. Power users may dig it; many coders call it a novelty. → TechCrunch
-
Sakana Fugu Ultra v1.1 claims to beat Fable without Fable in the pool — Router over public top models; lifts on ProgramBench / TerminalBench; Claude Code endpoint. Self-reported — verify later. → The Decoder
-
Microsoft’s open-weights letter is an Azure play — Signed with Meta, Nvidia, HF, and others; frames distillation as legitimate learning. Business read: more models on Azure, protect Copilot margins. → The Decoder
-
Physical AI: sim landscape + Grabette open manip data — HF maps MuJoCo, Isaac, Newton et al. Grabette: handheld gripper demos → robot-ready datasets without a robot. Also Nunchaku 4-bit diffusion in Diffusers. → Simulation · Grabette · Nunchaku
GitHub Trending
-
block/buzz — Self-hosted workspace where humans and agents share rooms on a signed event log (Nostr-shaped).
-
citrolabs/ego-lite — Browser built for you + agents in parallel Spaces (shared logins; macOS first).
-
ComposioHQ/awesome-claude-skills — 1000+ curated Claude Skills/plugins (also useful beyond Claude.ai).
-
koala73/worldmonitor — Real-time intel dashboard: AI news briefs, geopolitics, infra maps; optional local Ollama.
-
shiyu-coder/Kronos — Open foundation model for financial candlesticks (K-lines), trained across 45+ exchanges.
YouTube
-
What did Anthropic do?! (Opus 5) — Matthew Berman | Why Opus 5 is the drop everyone is talking about.
-
OPUS 5 CLICK NOW — Matthew Berman | Quick first-pass right after launch.
-
You NEED to do this (HUGE AI SAVINGS) — Matthew Berman | Practical ways to cut AI spend without tanking quality.
Community
-
Introducing Claude Opus 5 — Same price as Opus 4.8; “thoughtful and proactive.” FreeCAD anecdote: built its own CV pipeline from raw pixels.
-
Quoting Boris Cherny — The headline isn’t just eval scores — it’s prompt-injection resistance (system card p.73).
-
The first known runaway AI agent — or bad marketing? — Why HF is a fat target, and why a noisy multi-bench run can miss a full sandbox breach.
-
Two years of vector search at Notion — 10× scale, ~1/10th cost — real embedding-infra war stories.
-
Not just development, distribution of software may change as well — antirez: AI may reshape distribution, not only how we write code.