About the society
The Society for AI Security is an independent, practitioner-run knowledge base for AI security. We exist because the field moves faster than any one team can follow, and because no single institution keeps the whole record: the attacks, the defences, the tools, the incidents, the research and the rules, together and in the open.
Why a society
AI security is assembled from pieces held by different people. Application security teams own prompt injection. Cloud teams own the sandboxes agents escape from. Data teams own the training pipelines that get poisoned. Alignment researchers understand why a model pursues a goal past a boundary. Lawyers read Article 15, the Cyber Resilience Act, SB 53 and the RAISE Act. Standards bodies, safety institutes and the frontier developers publish frameworks, post-mortems and evaluations that few practitioners have time to read. The record that connects them did not exist, so members built it.
What we do
We maintain one public knowledge base: an incident database written for the mechanism and the missing control, an open source directory assessed for what each tool actually covers, a research library summarised for practitioners, a standards library that maps attack taxonomies to control catalogues, a regulation tracker read for security duties and reporting clocks, and a case study collection of programmes that were built. Long-form articles turn the record into controls, and a news service says what changed and what to do. Every record is written by a practitioner, reviewed by a second, dated and sourced.
Who we are
Members are security engineers and architects, AI red teamers and evaluators, CISOs, ML engineers, alignment and interpretability researchers, governance and risk leads, counsel, policy analysts and regulators. The society is governed by an elected council and working groups on agent security, red teaming and evaluation, regulation mapping, incident learning, and model and supply chain integrity. Membership is free for everyone.
What we are not
We are not a regulator, a certification body, a vendor or a law firm. We do not publish working exploits, jailbreak strings or attack payloads; we describe attack classes at the level a defender needs to recognise and test for them. Nothing on this platform constitutes legal or security advice for a specific system. We publish evidence so that the people responsible for systems can use it.