Announcements, product releases, and media coverage from ThirdKey. For interviews, briefings, or background on our work in zero-trust AI agent infrastructure, reach the press desk directly — we typically respond within one business day.
Product launches, specification publications, and milestones. Older items are archived but stay reachable by DOI.
Empirical study of steganographic exfiltration in vector databases. Attackers with write access to RAG pipelines can hide payloads inside embeddings (up to 1,920 bytes per vector at d=3072) using rotation, noise, scaling, and fragmentation that preserve retrieval behavior and evade statistical detectors. Introduces VectorPin — Ed25519 signatures binding embeddings to source and model. DOI 10.5281/zenodo.20076420 · arXiv:2605.13764.
3,760-trial adversarial evaluation across seven leading AI models: Symbiont reduces observed exploit success to 0% across filesystem, network, syscall, and state-mutation attacks — while industry-standard Docker sandboxing fails to stop 88% of filesystem and 84% of state-mutation attacks. Prompt-injection success cut from 59% to 3.6%.
A specification for .clad.toml declarative tool contracts — typed parameters, validation, invocation, output parsing, and Cedar policy across CLI, sessions, and governed browser execution. Replaces freeform shell generation with allow-list-validated tool invocation. DOI 10.5281/zenodo.19957595.
A compile-time approach to enforcing policy gates in AI agent loops via typestate encoding. Evaluated across nine hosted LLM providers — 263 forbidden tool-call attempts refused without execution at 30–95µs per check. DOI 10.5281/zenodo.19746723.
A system specification for zero-trust AI agent execution — declarative tool contracts, the Observe-Reason-Gate-Act loop, and structural separation of policy from model influence across five architectural layers. DOI 10.5281/zenodo.19636533.
For interviews, embargoed briefings, or technical background, email the press desk. Founder Jascha Wanger is available for comment on AI agent safety, policy-governed runtime, and supply-chain integrity.
We respond to journalists, analysts, and conference organizers within one business day. For sales or technical questions, please use sales@thirdkey.ai instead.
press@thirdkey.ai →ThirdKey is an AI safety and security company building open, interoperable trust infrastructure for autonomous AI agents — cryptographic identity, schema verification, declarative tool contracts, embedding-store integrity, and a policy-governed runtime. The company ships five open-source primitives (Symbiont, SchemaPin, AgentPin, ToolClad, VectorPin) and a commercial enterprise tier of its agent runtime for security teams running agents in production.
A guarantee that can’t be written down isn’t one.