Sunday, 25 October 2026

Worth catching up on

1. 

   Artificial intelligence

   Data-Void Exploits in Web-Searching LLMs

   A peer-reviewed security paper (to appear at NDSS 2027) shows that web-searching LLMs can be manipulated via 'data-void exploits': attackers publish disinformation in topic areas where reliable content is sparse, and the LLM's web search retrieves and cites that single adversarial page. Testing ChatGPT, Gemini and Sonar Pro across varied prompt tones and attacker strategies, the researchers found attacker content appeared in responses 31% of the time overall, with user skepticism offering only limited protection.

   arXiv cs.CRWed 21 Oct
2. 

   Artificial intelligence

   Stale Instruction Files Override the Code in Coding Agents: A One-Sentence Fix

   A new study tested Claude Code and Codex CLI across 40 Python projects where the project's instruction file (CLAUDE.md/AGENTS.md) specified a different convention than the existing code. Without guidance, agents followed the stale instruction file in 90.1% of choices, and Codex CLI's rewrites to match the file broke tests in 17 of 20 cases. Adding a single sentence telling the agent to prioritize the existing code over the instruction file reversed this, bringing code-following choices to 91.4% and test pass rate to 100%.

   arXiv cs.CRThu 22 Oct
3. 

   Artificial intelligence

   Amplifying Latency and Cost of LLM Inference via Rowhammer Attack

   Researchers demonstrate LLVMhammer, a Rowhammer-based attack that flips bits in Triton's JIT compilation pipeline to make it generate functionally correct but slower GPU kernels for LLM inference, with the degraded kernel persisting via Triton's disk cache. Tested across four 7-8B open models on NVIDIA and AMD GPUs, it raised inference latency 4-16% and, in some cases, corrupted outputs significantly (BERTScore as low as 25%).

   arXiv cs.CRThu 22 Oct