Research
Peer-reviewed research, co-authored by our co-founder and chief scientist, on finding where software drifts from its specification. It informs how JANUS reads, tests and monitors AI-powered software.
10
ACM SIGSOFT Distinguished Paper Awardsawarded to our chief scientist’s work
ICSE · FSE · ICPC · IEEE TSE
IEEE S&P · USENIX Security · NDSS · SIGIR
01 Specifications
Related research
JANUS Audit — SPEC Recovery
Checks real protocol implementations against their written specifications. Found 30 previously unknown inconsistencies across 10 implementations; 22 confirmed by developers.
Recovers how real apps are meant to behave from how they actually run, then tests it. Found 25 previously unknown functional bugs in 12 apps; existing tools found only 3 of them.
Maps each specification requirement to the code that implements it, in projects such as Nginx and OpenSSL. Found 11 previously unreported inconsistencies; 5 confirmed by developers.
Uses LLMs to write formal specifications from code. Produced verifiable specifications for 279 of 385 programs, ahead of Daikon and Houdini.
Infers the rules a system should always keep and turns them into runtime checks. Over 14% higher recall than prior methods, with almost zero false positives.
Interactive code maps that help auditors understand unfamiliar codebases; cut reliance on reading raw LLM answers by 79%.
ACM SIGSOFT Distinguished Paper Award
02 Vulnerability detection
Related research
JANUS Audit
Combines LLM reasoning with static analysis to find business-logic flaws that pattern-based tools miss. Found 9 new vulnerabilities missed by human auditors.
03 Model supply chain
Related research
JANUS Audit — AI-BOM
Shows model files can carry malicious behaviour. Found 1,083 abusable TensorFlow APIs; none of the resulting attacks were caught by Hugging Face, TensorFlow Hub or ModelScan.
Recovers the full structure and parameters of compiled models with millions of parameters, such as ResNet.
04 Adversarial testing
Related research
JANUS Pentest
Stops web search results from overriding an LLM's trusted knowledge. Repair success above 70% across mainstream LLMs.
Tests whether safety filters hold up against reworded prompts. Bypassed 5 safety checkers 54%+ more often than prior methods.
Uses teams of LLM agents to simulate 15 types of real-world insider attacks.
05 Runtime
Related research
JANUS Runtime MonitorEarly Access
Shows attack detectors lose up to 34% performance when “normal” behaviour drifts, and how to adapt.
Selected papers. Titles link to the publisher, the conference page or a preprint. Listed work is related research that informs JANUS; it is not product documentation.
or write to support@agentsure.tech