Six months after Anthropic's 'too dangerous to release' Mythos, the predicted vulnerability explosion never came
redbaron_4 · reddit · 2026-10-02
In April, Anthropic claimed its Mythos preview model was so good at finding and exploiting code vulnerabilities that it wouldn't be released publicly — instead shared with select companies under "Project Glasswing" to audit closed-source code, alongside disclosures including a long-standing FreeBSD networking bug.
Half a year later, with Mythos shipped and Fable following, the predicted vulnerability explosion hasn't materialized: most bugs disclosed since were in obscure options, and the AI-assisted Linux LPEs weren't Mythos-specific.
The author asks skeptically whether Glasswing produced any worthwhile contributions, suggesting Anthropic may have instead used otherwise-inaccessible closed source to train Fable.
More from Models
- MiniMax's M3.1 Flash joins RSIArena mid-run as the competition heats up — my_cat_can_code · 2026-10-02
- Why ChatGPT keeps answering 47 or 73 when asked for a number — Princevora03 · 2026-10-02
- Daily brief: Grok 4.7 rolls out as base model everywhere, Claude Code ships Mods — testingcatalog · 2026-10-02
- Screen Studio trained on 5,000 fake macOS apps to build a UI-specific upscaler — CurieuxExplorer · 2026-10-02
- Intelligence is jagged: use the cheapest, lowest-latency model that's good enough — sachdh · 2026-10-02
- Gemini 4 may roll out to Ultra users next week as Google's TPU capacity reportedly runs tight — haider1 · 2026-10-02