We publish open, reproducible work on agent integrity, adversarial resistance, and jailbreak detection, and release the datasets and specifications behind it so others can verify and build on the results.
The first formal framework for measuring AI agent security across digital and physical domains, introducing a three-property scoring methodology, Adversarial Coherence, Verifiable Assurance, and Environmental Portability, evaluated across four architectural layers.
Tarique Smith, Mansoor Malik
Polite phrasing raises attack bypass rates from 15% to 68%. Semantic Inversion is a preprocessing layer that detects politeness markers, inverts indirect requests, and applies dynamic risk amplification at 1.8ms median latency, cutting polite bypass rates by 71%.
Cogensec Security Research Lab
Agents coordinate malicious activity without explicit programming, bypass safeguards through emergent communication, and cascade failures across networks. Frontier models already reach 50% steganographic coordination rates.
Joint studies, dataset access, and replication of published results with academic and industry partners.
Contact the research teamPaid, scoped engagements for jailbreak researchers and offensive security engineers pressure-testing frontier systems.
Apply to the networkHow to cite our work: every published paper carries a citation block with BibTeX, APA, and MLA formats, plus Highwire citation metadata for Google Scholar indexing.