LLM & Agent Security and Privacy.

Security, privacy, safety, provenance and abuse of language models, AI agents, multimodal models, text-to-image systems and diffusion models. This includes prompt injection, jailbreaks, poisoning, extraction, inference attacks, private computation, watermarking, unlearning, and relevant human studies. Selected general ML privacy-auditing, machine-unlearning, and NLP/security papers are labelled as foundations. General ML security is not exhaustively covered.

Official main-conference listings define the source inventory. Retrieved abstracts are screened for scope, followed by a review of candidates and boundary cases. Every record retains its source and review status. Missing abstracts stay unresolved. This is an assisted research catalogue, not a claim that every full paper has been independently read.

Posters, demos, workshops, invited talks and papers that only use AI to solve an unrelated security problem. Work on attacks against AI security tools, model misuse, and reliability or privacy of AI assistance can be included. Conventional voice assistants and standalone audio generation are outside the current scope.

A monthly GitHub workflow checks official conference listings for new titles and possible new editions. A monthly cloud agent retrieves abstracts, retries missing evidence, screens papers and proposes catalogue updates in draft pull requests. New conference editions get separate setup PRs; screening begins after you merge their setup. Reviewed changes are published after their PR is merged. Listing checks, abstract retrieval and scope review are tracked separately.

Years refer to conference editions, not preprint dates. CCS 2026 papers are accepted and forthcoming as of October 23, 2026. Listings, metadata and publication status can change.