# Where was Mythos when WordPress fell? > [!info] Ported from an existing deep-read (Refactor Appendix present). Written with a disclosed conflict of interest (Claude/Anthropic generated the original analysis of an article critiquing Anthropic) — treat critical judgments in the source deep-read with that discount already applied. Several claims from this source were already promoted to Permanent notes on 2026-08-08, before this Literature Note existed; this note fills that gap and connects rather than re-derives. > [!question] Lingering Questions > - Which thousand projects did Mythos Preview actually scan? Unanswerable except by Anthropic — the article's central unresolved question. > - What's the baseline triage-survival rate for human-submitted vulnerability reports or conventional-scanner output? Without it, the article's own "~60% held up" figure can't be scored as good or bad — its biggest analytical gap. > - Has anyone attempted to reproduce Kues' wp2shell methodology (stripped git history, pinned environment, capped agents) under matched conditions? > - Would the same ~$25/10-hour method find comparable vulnerabilities in the top 100 WordPress plugins by install count? Nobody should publish the answer, but someone defensive should know it. > [!tip] Article Ideas > - "A model grading its own homework" — self-assessment vs. independent verification, using this article's own numbers (23,019 total findings / 6,202 self-rated high-critical / 1,752 independently checked) as the worked example. Cross-link the Grafana vulnerability-response talk from the same batch — maintainer overload from the receiving end. > - Steering vs. capability — the article raises and abandons this question; it's the most underexplored thread here and already has vault traction (see harness-design notes below). > [!link] Cross Reference > - [[3. Resources/Permanent Notes/A Vendors Absence From a Credits List Is Weaker Evidence Than It Sounds]] — already promoted from this exact source (2026-08-08); covers the article's argument-from-absence weakness > - [[3. Resources/Permanent Notes/Maintainer Attention Is the Scarce Resource in Open Source]] — covers the ~530-reported/75-patched maintainer-overload claim > - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — covers the self-rated-severity claim > - [[3. Resources/Permanent Notes/The Harness Now Outpaces the Model as the Source of Agent Capability Gains]] and [[3. Resources/Permanent Notes/Harness Design Can Matter More Than Model Choice for Agent Performance]] — cover the steering-vs-capability thread > - [[3. Resources/Permanent Notes/First-Principles Source Reading Finds Bugs Patch-Diffing Hides]] — from Kues' own wp2shell writeup; covers the "remove the answer key" methodology this article reports on secondhand > - Courtney's own `~/.claude/CLAUDE.md` "Agent code-analysis / audit tasks" section — directly derived from the Kues prompt this article describes; the guardrails there currently carry no written provenance back to this piece > [!quote] Vocabulary > - **wp2shell** — Adam Kues' (Searchlight Cyber) name for the pre-auth WordPress core RCE exploit chain he found by pointing an agent at a git-history-stripped WordPress tree > - **Project Glasswing** — the program name under which Mythos Preview was released to trusted entities ## Bibliographic capsule Personal blog (patrik.re), ~1,079 words, published 2026-08-05. Sceptical audit of the gap between Anthropic's published claims for Mythos Preview and the observable public record, built almost entirely from the vendor's own disclosed figures plus linked independent accounts (Stenberg on curl, Schneier's commentary, Cloudflare's operational writeup). Category: article. ## Author's thesis (as synthesized) A security model marketed as "securing critical software for the AI era" did not surface the most consequential WordPress vulnerability of the period — a pre-auth core RCE found instead by a human researcher using a different model and a specific methodology — and the vendor's own published numbers, read carefully, show most of the headline finding count is self-assessed rather than independently verified. ## Key arguments (≤3) 1. **The self-assessment gap**: of 6,202 findings self-rated high/critical, only 1,752 were independently checked (~28%), and about 60% of that checked subset held up — a legitimate methodological objection built from arithmetic on the vendor's own disclosed numbers, though the 60% figure itself is presented with no comparison baseline. 2. **The externality**: ~530 high/critical reports landed on volunteer open-source maintainers, 75 were patched, and several maintainers asked Anthropic to slow the disclosure pace down — automated discovery at scale converted into unfunded maintainer labor. 3. **The methodological counter-example**: Adam Kues found the actual WordPress core RCE by stripping git history from a clean WordPress tree (so an agent couldn't shortcut by reading patches), capping agent count, and letting it run roughly ten hours for about $25 — a reproducible methodology, not a lucky break, and the direct ancestor of Courtney's own CLAUDE.md audit guardrails. ## Claim atoms ### A model grading its own homework > [!quote] ID: rw_01kze6zd3wx523cc7ypqm0514n > "the number that kept popping up in my feed was a model grading its own homework, and the part where somebody else marked the homework covered about a quarter of it" Where/why: the article's strongest section — simple, checkable arithmetic on Anthropic's own published figures. Already promoted; see [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]]. ### Securing critical software arrived as a ticket queue for unpaid volunteers > [!quote] ID: rw_01kze6zd3wx523cc7ypqm0514n > "several maintainers told them they were so overloaded that they asked Anthropic to slow the disclosures down" Where/why: the article's most humane finding — scaling vulnerability *discovery* without scaling *triage capacity* externalizes cost onto the people least equipped to absorb it. Already promoted; see [[3. Resources/Permanent Notes/Maintainer Attention Is the Scarce Resource in Open Source]]. ### Remove the answer key Where/why: Kues stripped git history from a WordPress tree specifically so the agent couldn't shortcut by pattern-matching known patches, then capped agent fan-out and let it run. Methodology, not luck — and reproducible in principle, though nobody has attempted to reproduce it yet. ### 60% held up, with no baseline to compare it to Where/why: presented as damning, but human-submitted vulnerability reports and conventional-scanner output also fail independent triage at some rate the article never states. This is the deep-read's own flagged "biggest analytical gap" — a verified-fraction number that can't be scored as good or bad without a denominator to compare it against. Not yet captured by an existing Permanent (see below). ## Permanent candidates - Self-grading, maintainer-overload, argument-from-absence, and harness/steering claims were **already promoted from this source on 2026-08-08** — see Cross Reference above. Connected, not duplicated. - **A verified fraction without a comparison baseline can't be scored** — genuinely uncaptured elsewhere in the vault; promoted below. ## Processing checklist - [x] Checked for an existing Literature Note on this doc (none found) - [x] Checked existing Permanents before creating new ones — found five already covering this source's other major claims, connected rather than duplicated - [x] Connected into [[3. Resources/MOCs/Open Source Governance MOC]] and [[3. Resources/MOCs/AI Agent Safety MOC]] - [ ] Fetch and independently verify the linked primaries (Anthropic's 2026-05-22 update figures, Stenberg's curl evaluation, WordPress 7.0.2 credits) before citing any number from this piece elsewhere — the deep-read explicitly did not do this ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] — self-verification failure and the maker/checker gap in vendor-reported AI security findings - [[3. Resources/MOCs/Open Source Governance MOC]] — maintainer-overload externality from automated vulnerability discovery at scale - [[3. Resources/Readwise/Articles/Where Was Mythos When WordPress Fell]] — source article