self.md radar — 2026-05-08
Agents crossed into places where rollback is expensive.
Google supplied the silicon-grade proof. Mozilla and HiddenLayer supplied security evidence from both directions. The workflow repos had the boring answer: rails, logs, and stop conditions before the agent writes.
1. AlphaEvolve reached silicon-grade review
sources:
what happened: Google DeepMind says AlphaEvolve has moved from pilot testing to a core infrastructure component. It is now a regular tool for next-generation TPU design. DeepMind says it found cache-replacement policies in two days that previously took months, helped reduce Spanner write amplification by 20%, and produced compiler ideas that cut software storage footprint by nearly 9%. Jeff Dean’s quoted example is the sharp one: a counterintuitive circuit design went into next-generation TPU silicon.
why this matters: Review moves below app code. A suggestion can become a cache policy, compiler strategy, or circuit path, so the hard part is proving the rule before it ships into hardware or storage.
2. Security agents became both scanner and attack surface
sources:
what happened: Mozilla published the details behind its Claude Mythos Preview run: 271 bugs identified by Mythos Preview, 423 total security bugs fixed across April, and Firefox 150 internal rollups containing 154, 55, and 107 bugs. In the same window, HiddenLayer found a Hugging Face repo called Open-OSS/privacy-filter with more than 200k likely inflated downloads, a copied OpenAI model card, and loader.py/start.bat paths that pulled a JSONKeeper command into hidden PowerShell. GitHub’s Claude Code advisory adds the local-agent version: CVE-2026-39861, patched in 2.1.64, let symlink following write outside the workspace after untrusted context triggered sandboxed execution.
why this matters: Same story, two doors. Agents can find bugs humans miss, and agent-adjacent repos can put malware or file writes exactly where credentials live. The boundary that matters is which process can write where, and which repo a tired operator runs.
3. Skills made prompts enforceable
sources:
what happened: Addy Osmani’s agent-skills repo packages seven slash commands and 20 skills around spec, plan, build, test, review, simplify, and ship. The useful pieces are the machinery around the prompt: verification gates, anti-rationalization tables, and hooks that make skipping the boring step harder. Vercel’s Open Agents repo pushes the same idea into runtime shape with a web-to-workflow-to-sandbox-VM stack, durable workflow runs, snapshot-resumable sandboxes, and optional auto-commit / auto-PR. GitHub’s Spec Kit adds constitution and spec commands before implementation starts.
why this matters: Prompt work is becoming procedure. The artifact has to say what counts as done, what blocks a commit, and what evidence survives after the sandbox closes.
supporting links
- Natural Language Autoencoders — Anthropic’s activation-to-text method is useful partly because it admits ugly limits: hallucinated explanations and hundreds of generated tokens for every activation it reads.
- AgentGuard / ConfigGuard — a browser-local scanner for AGENTS.md, Cursor rules, MCP configs, and Codex-style settings; basically lint for the files that grant agents authority.
- ok-skills — another skills/playbooks repo for Codex, Claude Code, Cursor, and OpenClaw, which makes the instruction-packaging pattern harder to dismiss as one repo having a moment.
left on the table
- DeepSeek-TUI was an exact seen URL from 2026-05-05, so it stayed out.
- Gemma MTP for llama.cpp was too close to the 2026-05-06 Gemma MTP support note.
- Mozilla’s Mythos writeup repeats an actor and theme from recent radar, but it stayed because the new post adds official bug counts and CVE rollup detail.
- South Korea’s Buddhist robot thread had cultural weirdness, not enough source weight.