Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours
Source: The Decoder (opens in a new tab) · Matthias Bastian
Intel Summary
Epoch AI researchers reported that the open-weight AI model GLM 5.3 autonomously generated full exploit chains targeting six Firefox releases by analyzing public security fixes. Operating with zero human input in under 48 hours, the model developed functional exploits before updates reached users, costing approximately $1,000 in API fees per release without triggering safety filters from tested providers.
Why It Matters
Autonomous reverse-engineering of public patches dramatically compresses the defensive window between code fix disclosure and user-side software updates. This capability demonstrates that low-cost, automated offensive tooling can weaponize public patch diffs before distributions deploy, complicating traditional vulnerability disclosure practices and exposing gaps in API guardrails.
Organizations & Entities
Related Intelligence
- DevelopmentNewAlso involving Epoch AI
Epoch AI reports GLM 5.3 autonomously generated Firefox exploit chains
Epoch AI researchers reported that the open-weight AI model GLM 5.3 autonomously generated full exploit chains targeting six Firefox releases by analyzing public security fixes. Operating with zero human input in under 48 hours, the model developed functional exploits before updates reached users, costing approximately $1,000 in API fees per release without triggering safety filters from tested providers. Claims are as reported; this summary makes no determination about accuracy or significance.
- ReportAlso involving GLM-5.3
Paris-based Mistral releases "Le Chonk," a strong coding model that still trails China overall
Mistral has released the open weights for Large 4, an open-weight, trillion-parameter model trained entirely in Europe nicknamed "Le Chonk." Mistral claims the model outperforms open competitors at coding tasks, while independent testing indicates it operates quickly and at relatively low cost. However, it trails leading Chinese models, posting a score of 26 compared to 32 from DeepSeek V4.1 Flash and GLM-5.3.
The Decoder - Report
Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity
Google has launched Gemini 4 Carbon, its latest frontier model claiming substantial improvements in cybersecurity, coding, and agentic workflows. According to Google, the model ranked first on the Artificial Analysis Intelligence Index (v5.0) benchmark, outpacing frontier models including Claude Opus 5.5 and GPT-6 Astra.
The Verge - Report
‘Nobody’s Ready’: AI-Found Bugs Spark a Corporate Security Spending Spree
Enterprise software startup Atomicwork directed its engineering team to halt standard work and focus entirely on cybersecurity following multiple incidents over a 48-hour period. Reported compromises included a breach at root certificate authority eMudhra, a security compromise involving open-source AI request routing tool LiteLLM, and a newly discovered vulnerability in PHP.
The Information