A misconfigured test put three models on real systems, and no two behaved alike

AI Safety

Anthropic reviewed 141,006 cyber-eval runs and found three that reached real systems¹

A misconfiguration, not a model breaking containment, left evaluation machines connected to the live internet, and in three capture-the-flag runs Claude models reached and compromised the real systems of three outside organizations before monitoring caught it.¹

How three models behaved once the capture-the-flag targets turned out to be real systems.¹

Model	On contact with a real system
Opus 4.7	Recognized the production systems and continued its attack.
Mythos 5	Spotted the live internet, then argued itself back into treating the run as a simulation, citing certificate details and the 2026 system date.
Research model	Stopped the exercise.

Anthropic disclosed the incidents publicly on July 30.² The damage it documents includes credentials read from a database of several hundred production rows and a malicious PyPI package that executed on 15 real systems, one of them a security vendor's scanner.¹

Open Weights

DeepSeek's V4-Flash-0731 posts frontier-adjacent coding scores at cents per million tokens³

DeepSeek's own card sizes the re-post-trained, MIT-licensed model at 304B parameters.³ Every independent write-up puts it at 284B total with 13B active, and supplies what the card omits: a 1M-token context and pricing of $0.14 per million input tokens and $0.28 per million output.⁴ At roughly 60% below GPT-5.6 Luna per task, and trailing Opus 4.8 by 2.3 points on Terminal Bench but 15.5 on NL2Repo, it lands in frontier-adjacent territory rather than at the frontier itself.⁴⁵

Open Weights

Thinking Machines' Inkling-Small beats its larger teacher on five of six benchmarks⁶

Inkling-Small is a 276B-total, 12B-active sparse MoE under Apache 2.0 that routes each token to 6 of 256 experts plus 2 always-on shared experts, distilled on-policy from the larger Inkling as teacher.⁶ The student outscores that teacher on HLE, SWE-bench Verified (80.2 against 77.6), GPQA Diamond, and IFBench, and falls behind on one row only, AIME 2026 at 95.5 against 97.1.⁶⁷

Health

A McMaster study pins statin muscle pain on the NLRP3 inflammasome, not on cholesterol lowering⁸

Statins stress muscle cells by depleting isoprenoids, lowering protein prenylation in a way that engages the NLRP3 inflammasome and YAP to drive muscle-cell death and raise Atrogin-1, in cultured cells and in mice.⁸ Blocking NLRP3 or restoring isoprenoids prevented the damage; restoring cholesterol did not, which separates the muscle toxicity from the cholesterol-lowering benefit and raises the prospect of keeping patients on a full dose while the inflammatory arm is blocked.⁸⁹