The National Academies Launders Mythos: \u201cImplications of AI for Cybersecurity\u201d<\/a><\/p>\n https:\/\/www.flyingpenguin.com\/the-national-academies-launders-mythos-implications-of-ai-for-cybersecurity\/<\/a><\/p>\n Publish Date: 2026-06-26 03:16:00<\/a><\/p>\n
| NASEM Laundry (June 2026)<\/th>\n | Prior Evidence<\/th>\n<\/tr>\n |
|---|---|
| Figure 1 plots Mythos at 83.1% on CyberGym as settled capability, sourced to \u201cWang et al. 2025\u201d<\/td>\n | The 83.1% has been repeatedly proven false. It\u2019s a self-reported number by Anthropic. AISLE proved detection reproduced in 8 of 8 open-weight models, even at $0.11 per million tokens, Cisco proved outcome is model-independent<\/td>\n<\/tr>\n |
| Restricted Glasswing access presented as responsible handling of uniquely capable model<\/td>\n | The danger warnings are self-serving FUD marketing. Model uniqueness repeatedly disproven. Mythos emailed out of its sandbox only after being instructed to try, showed no sign of altering its weights, and Opus 4.6 finds the same or better flaws<\/td>\n<\/tr>\n |
| Vulnerability discovery framed as a breakthrough enabling novel risk<\/td>\n | The flagship FreeBSD CVE-2026-4747 is a 2007 patch in training data, opposite of novel. It was a curated recovery from a backlog of delayed fixes, which any model does.<\/td>\n<\/tr>\n |
| Benchmark score offered as capability evidence<\/td>\n | Of 23,019 reported findings, 1,752…<\/td>\n<\/tr>\n<\/table>\n |