Tap Notes: Nobody Reads the Diff

Most of today’s reading is about the same failure: a gate that exists on paper and doesn’t work in practice. The approval prompt gets clicked through. The scanner’s output lands on a maintainer who has to verify it. The agent runs for 35 hours and nobody could review what it wrote. “Human in the loop” is a diagram, and the human has other things to do.
The Gates
After 7 months, Glasswing overwhelms maintainers with underwhelming findings VulnCheck audited seven months of Anthropic’s Glasswing scanner: about 38k findings, 217 fixed, and curl rejecting more than half of the reports filed against it.
Why it matters: A “94% true positive” claim is cheap when someone else does the validating. The cost didn’t disappear, it moved to the people least able to absorb it. If you ship AI-found bugs upstream, the unit of quality is how much maintainer time each report costs, not how many reports you filed.
Auto mode is now the default in Claude Code for Pro, Max, and Team plans (from August 2026) Anthropic made auto mode the default and backed it with evals. In one test of 1,053 paid testers, 86.4% approved a clearly dangerous swapped command. Auto mode would have blocked most of those. Simon Willison accepted the comparison but wanted independent confirmation.
Why it matters: That 86.4% is the cleanest measurement of confirmation fatigue I’ve seen. It reframes the question from “is the classifier safe?” to “safer than what?” The answer is humans rubber-stamping everything. But Simon’s residual case still stands. A malicious package whose setup instructions read as normal isn’t injection, it’s supply chain, and no permission classifier sees it. Trust the classifier a little. Build as if it weren’t there: least privilege, scoped credentials, closed egress.
Stealing Reasoning Traces from Proprietary LLM APIs (from August 2026) Researchers showed that encrypted reasoning blocks from the major providers could be replayed across sessions, and that models in one family shared a key. Replay a trace into a weaker sibling, jailbreak it, and the hidden reasoning comes out in plaintext. The providers say the attacks no longer work.
Why it matters: The key-sharing flaw is fixed. The injection variant in the paper is the durable part. Models trust their own reasoning, so anything smuggled into a trace gets treated as self-generated intent. If your agent round-trips thinking blocks, that’s a trust boundary you didn’t draw.
The Output Nobody Can Review
Astra for Coding: Why Are We Doing This Again? (from September 2026) Armin Ronacher ran an unattended GPT 6 “Astra” software factory for 35 hours: roughly 4B tokens, $1,200, 75k lines of code, nothing of value. His argument is that training rewards token efficiency and long-horizon completion, so the model code-golfs everything, including what gets committed.
Why it matters: Readability has no cheap proxy, so the optimizer drops it. Then it runs unattended and commits its own reward hacks. His task-name drift from 1, 2, 3 to 8b2c2b3 checkpoint1 is a nice record of an agent losing the thread. You can’t lint your way out of this. The fix is in the harness: forced edit tools, diff-size budgets, a human-readable bar for merges.
Prompting Claude Fable 5.2 Anthropic’s guide pairs observed behaviors with tested prompt text and measured effects. The “say who owns the code” section is the sleeper: one sentence moves the model from appending politely to merging and fixing unprompted, and a “no drive-by refactors” rule doesn’t cancel it.
Why it matters: That’s a hidden dial for anyone writing agent harnesses. Two other findings are worth keeping. Cost-first subagent descriptions cut delegation by about a third with the same success rate. And caution accumulates in carried-over context. Old “never do X” notes silently neuter later sessions, so memory rules need a reason and a reviewer, not just an append.
The Quiet Fix
Show HN: Get your coding agent off ~2019 JavaScript, onto modern web primitives The Chrome team shipped 200+ curated guides as an agent skill with a small local search CLI. Models know modern platform features exist but default to the old JS pattern. In their evals across three agents on 200 tasks, pass rates went from 55% unguided to 92% guided.
Why it matters: Receipts for “skills work.” The interesting engineering is the retrieval: they shrank a 500MB native stack to a ~7MB pure-JS package with ~10ms search. That keeps the guidance out of your context window until it’s needed. They’re also using the evals to prune guides once models learn the thing on their own, which is how a skill library should age.
One more thing
The Single Space That Almost Erased My Personality (from August 2026) A local assistant answered as a generic bot for dozens of turns because a placeholder prompt beat the persona. The obvious fix, letting an explicit override replace the persona, would have broken a caller that deliberately passes a single space to satisfy a truthiness check.
The real fix made the persona the base and added a proper boolean opt-out. As a side effect, confabulated capability claims went away, because the persona carried the honesty guard the placeholder didn’t. Voice and honesty were the same missing layer. A check that runs once, at the wrong layer, isn’t a check. 🪨