Tags
When started with --reranking, llama.cpp lets a remote attacker trigger a denial of service (std::bad_alloc, HTTP 500) via a negative top_n on POST /rerank (CVE-2026-52132).
Read full articleA crafted GGUF file with an empty metadata key reaches an assertion in llama.cpp's gguf_reader::read and aborts the process. Affects any binary that loads GGUF files (CVE-2026-52131).
Read full articleA deeply nested JSON schema exhausts the recursion stack in llama.cpp's grammar converter, crashing the server. Only POST /completions is affected (CVE-2026-52130).
Read full articleMultiple studies conclude that an LLM agent's completion bias and overconfidence cannot be corrected by prompting. This post lays out the design principles of a vulnerability-checking automation harness that makes "surveyed everything" and "0 vulnerabilities" computed from an HMAC receipt ledger and deterministic hooks rather than from the model's narrative. A "200 OK" without a negative control is not confirmation.
Read full articleIn an era where LLM agents have started solving CTFs automatically, how do you design a challenge that "machines struggle with but a skilled human solves in a reasonable time"? This organizes the anti-LLM design principles that target the structural limits of transformers — state tracking, carry propagation, tokenization, and context loss.
Read full article