Review batch 8: PIR-2026-0041..0045 corrected on adversarial review

[srinivasgumdelli]

Eighth batch of the adversarial review loop (Codex, an OpenAI model; affiliation noted in every dated note). Primary re-read for this PR: Censys's OpenClaw exposure post.

• 0042 OpenClaw exposure: Censys establishes the adoption wave, 21,639 publicly accessible instances on 2026-01-31, the local-binding and SSH-tunnel guidance, and that most exposed instances still required a token. It does not establish credential or history leaks; those are now attributed to later studies and coverage on the vulnerable subset, with the 42,665 and ~5,194 figures marked as the later study's. The CVE stays a separate researcher-demonstrated component.
• 0041 ClawHavoc: "delivered, or were built to deliver" AMOS in the title and mechanism, since the artifacts prove capability and no victim telemetry proves execution; the 341 / 824 / later counts kept with denominators and dates; controls_that_worked separates post-discovery controls from preventive ones.
• 0045 SEED key leak: direct_loss_usd from 0 to unknown; the $30 cap bounded exposure but a zero was not audited.
• 0044 Grok/Bankrbot and 0043 Lobstar Wilde: no material change; dated review notes added.

INDEX: three rows updated; enum counts unchanged.

🤖 Generated with Claude Code