网络安全趋势综述：ValsAI 称 V4.1 Flash 在模拟攻击任务上胜过 GPT-6 Sol

Michael Tomlinson @mikejt4原推发布于 2026.10.24

在 X 上打开原推

Where this cyber wave is heading My post this morning listed a fortnight of exploited bugs. Zooming out, four threads keep showing up in the primary sources. 1. AI on both sides 1. DARPA's AI Cyber Challenge (Aug 2025): autonomous systems found 54 of 63 planted bugs in 54 million lines of code and patched 43, plus 18 real zero-days. 2. Google's Big Sleep agent found a SQLite flaw only attackers knew about and got it fixed before it was used. 3. Mandiant's M-Trends 2026 puts mean time-to-exploit at -7 days, i.e. exploitation is often seen before a patch exists. 4. Anthropic mapped 832 banned accounts: the share of medium or high risk actors went from 33% to 56% in a year. 5. Fortinet's hardening release this week credits frontier-AI-assisted testing (Anthropic, OpenAI, Google, SpaceX) for bugs it fixed. 2. Open weights: both edges @joshua_saxe argues open-weight cyber models are currently net good for defence - cheaper tokens, on-premise, air-gapped, no secrets leaving the building. @AndrewCurran_ flagged defenders mid-attack getting refusals from closed models. The other edge: @ValsAI found DeepSeek V4.1 Flash beat GPT-6 Sol on their simulated network attack task, and community "abliterated" builds with the refusals stripped out are free to download on Hugging Face. Moonshot says it cut cyber training data from Kimi K3.1 Pro before releasing weights, and its exploit score still only fell from 90.5% to 81%. 3. Alignment UK AISI says the length of cyber tasks models can finish on their own was doubling every 4.7 months, and the newest models beat that trend. Anthropic has published an assessment of four incidents where Claude models got unauthorised access to real third-party systems during testing. As agents get more autonomy, a model that misreads its instructions becomes a security incident in itself. xAI says Grok 4.7 lets only 3.3% of risky dual-use prompts through while rarely blocking legitimate security work, and is giving select defenders red-team access. Credit where it's due!! 4. Next 1-3 years (my speculation) 1. Patch windows go from weeks to days, so auto-updates stop being optional. 2. Supply chain becomes the main battleground - LiteLLM's hijacked release and the eMudhra certificate keys both hit things we trust by default. 3. Every serious vendor runs AI bug-hunters on its own code before attackers do. 4. Smaller shops lean on managed services and passkeys because they can't staff a 24/7 team. Sources https://www.darpa.mil/news/2025/aixcc-results https://blog.google/innovation-and-ai/technology/safety-security/cybersecurity-updates-summer-2025/ https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026/ https://www.anthropic.com/research/attack-navigator https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents https://x.ai/news/grok-4-7 https://www.kimi.ai/blog/kimi-k3-1-pro https://docs.litellm.ai/blog/security-incident-oct-21-2026