Provenance Record

Well-structured News piece with a clear Overview / What We Know / What We Don't Know format. 782 words, within the 400-1200 News range. Every bullet in 'What We Know' is individually attributed to a specific outlet, and the article is careful to distinguish figures that come from different tests (e.g. the 84%/13% figure vs. the separate 90%/23% comparison) rather than conflating them.

All 4 source snapshots read in full from local gzip snapshots (sha256 of decompressed content verified against manifest.json for all 4 files — all matched exactly, all fetched with HTTP 200, no archive fallback needed). (1) sources/2026-07/.../source-0.html.gz (MIT Technology Review, 'Meet GPT-Red...') confirms: GPT-Red as internal LLM attacker built via self-play, the fake-chain-of-thought discovery, the Choquette-Choo '1+1=3' quote verbatim, the Hunn quotes verbatim, the Kandpal 'risk surface/blast radius' quote verbatim, the Vendy/Andon Labs test, the GPT-5 vs GPT-5.6 90%/23% figures with GPT-5 dated 'August last year' (=August 2025, correctly resolved by the article since this MIT TR piece published July 15, 2026), and that OpenAI will not release GPT-Red. (2) source-1.html.gz (SiliconANGLE) confirms: 84%/13% success-rate figure verbatim, the 'sixth of the rate...four months earlier' / 6x figure, the fake-chain-of-thought 95%->10% figure verbatim, the Hunn and Kandpal quotes (attributed to MIT Technology Review by SiliconANGLE itself, correctly reflected in the article), and the Vendy agent compromise. (3) source-2.html.gz (Decrypt) confirms: the self-play training description and the exact 'GPT-Red learns through adversarial self-play...' OpenAI quote, the 84%/13% figure, the Vendy case study (lowering prices, ordering discounted inventory, canceling an order) with vulnerabilities disclosed and addressed, and the 'GPT-Red will remain an internal tool because it contains intentionally developed offensive capabilities' line verbatim. (4) source-3.html.gz (Help Net Security) confirms: the GPT-5.1 84% indirect-injection figure, the Vendy real-world test with the exact $0.50 price-floor detail verbatim, the 6x-fewer-failures / 'GPT-5.6 Sol' framing verbatim, the fake-chain-of-thought 'below a tenth' quote verbatim, the 97% indirect-injection accuracy and 0.05% direct-injection failure figures verbatim, and the closing OpenAI quote about scaling compute for future GPT-Red versions verbatim. No hallucinated quotes, no misattributed specifics, and no orphan body URLs outside article.sources found across any of the four snapshots.

Every number, name, and direct quote in the body traces exactly to one of the four cited outlets. The article correctly flags in 'What We Don't Know' that the 84%/13% figure (SiliconANGLE/Decrypt) and the separate 90%/23% figure (MIT Technology Review) come from methodologically distinct tests rather than presenting them as a single consistent statistic — this is an accurate and appropriately cautious editorial call given the sources genuinely describe different test setups. No fabricated specifics found.

High-quality, thoroughly sourced submission. All four sources independently verified against every claim, quote, and statistic in the article with no discrepancies found. The sole automated finding is a source-allowlist configuration gap for an unambiguously reputable outlet, not a content defect, so APPROVE is the correct verdict rather than APPROVE_WITH_CORRECTIONS — there is nothing a corrections note could honestly tell readers to fix.