self.md radar — 2026-05-01
the important change today is that AI systems are being trusted with whole operational loops, not just one-off outputs.
three concrete systems land in this edition: an external cyber evaluation that treats a frontier model like an attack operator, a Claude Code-based personal stack that treats the assistant like a runtime instead of a chatbot, and a research agent that carried a physical optics program through thousands of steps instead of stopping at idea generation.
1. GPT-5.5 becomes the second model to finish the whole attack loop
sources:
what happened: AISI published its evaluation of GPT-5.5 and says it is the second frontier model on record to solve one of the institute’s multi-step cyber-attack simulations end-to-end. On the expert-level set, GPT-5.5 scored 71.4% versus 68.6% for Claude Mythos Preview, 52.4% for GPT-5.4, and 48.6% for Opus 4.7. In the rust_vm challenge, AISI clocks GPT-5.5 at 10 minutes 22 seconds and $1.73 on a task that took a human expert about 12 hours, run inside a basic ReAct scaffold with Bash and Python tools in a Kali Linux container. NCSC posted same-day guidance telling defenders to assume some attackers already have frontier AI tools and to put comparable systems on the defensive side of the line.
why this matters: This is an operator benchmark crossing a threshold, not a leaderboard nudge — once a single model can drive the full kill chain in a sandbox, the planning posture for both attack simulation and defensive tooling has to change.
2. PAI 5.0 packages Claude Code as a life OS runtime
sources:
what happened: Daniel Miessler released PAI 5.0 today as an open-source framework for running a “Life Operating System” on a personal machine. The stack is split into three layers — PAI as framework, Pulse as dashboard, and a named Digital Assistant as the user-facing personality — with Claude Code as the runtime, Bun as the toolchain, and TypeScript across the project. The operator pieces are concrete: Algorithm v6.3 governs multi-step task flow, three persistent memory surfaces hold state, deterministic hooks fire at Claude Code lifecycle events, specialist subagents take on scoped jobs, five inspectors watch every tool call, and skills self-activate when their trigger conditions match.
why this matters: Treating Claude Code as a runtime rather than a chatbot is a real architectural claim, and PAI is one of the first public stacks to wire memory, hooks, subagents, and inspectors into the kind of coherent execution surface that personal automation has been missing.
3. A lab agent claims a physical discovery, not just a literature summary
sources:
what happened: A new arXiv preprint introduces Qiushi Discovery Engine, an LLM-based system positioned for end-to-end autonomous scientific discovery on a real optical platform. The authors report that the agent reproduced a published transmission-matrix experiment on a non-original setup, translated coherence-order theory into experimental observables, and then proposed and experimentally validated an optical bilinear interaction the paper frames as analogous to a core operation in Transformer attention. The open-ended study consumed 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes, and 44 scripts. These are paper claims, not settled consensus.
why this matters: The interesting thread for self.md is duration: the agent stayed inside the measurement-and-revision loop long enough to surface a physically grounded result, which is a different shape of work than “agent proposes, human runs.”
supporting links
- Codex 0.128.0 — persisted /goal workflows and runtime continuation make long-running objectives explicit in the terminal.
- LLM 0.32a0 — prompts, reasoning, tool calls, and attachments are being serialized as first-class message parts.
- Step-level Optimization for Efficient Computer-use Agents — a fresh paper arguing the frontier model should wake up only at risky moments, not every GUI step.
left on the table
- Figure production metrics — strong, but yesterday’s published edition already used Figure as the proof that robotics was hitting real throughput.
- Japan Airlines humanoid deployment — same real-world robotics wave inside the 7-day window, and thinner sourcing than the signals we kept.
- Codex 0.128.0 — useful operator footnote, but too close to yesterday’s IDE / delegation edition to deserve a full standalone slot.
- allowed repeat: AISI on GPT-5.5 cyber capabilities — same GPT-5.5 family as the 2026-04-24 edition, but this one earned inclusion because it is a real external benchmark and data drop, not a recycled launch angle.