Antaeus: Mining Repository-Level Logic Vulnerabilities with Context-Based LLM Reasoning

The researchers propose the Antaeus framework, which detects logic vulnerabilities by anchoring LLM reasoning in repository-level code context; on 35 repositories containing confirmed logic vulnerabilities, it detected and explained 20 vulnerabilities, outperforming function-level and agent baselines at similar token usage and cost. The paper points out that identifying logic vulnerabilities requires inferring application-specific security invariants and implicit behavioral assumptions, which even frontier agent models struggle to do, because these invariants are often implicit and buried in unrelated code. Antaeus first ranks functions using lightweight repository-level security signals, concentrating expensive LLM analysis on relevant code to reduce the number of calls, cost, and triage workload; it then combines local code context with a repository-level view of application functionality, security resources, and trust boundaries to identify security-sensitive sinks, derive security execution conditions, and check whether they are satisfied locally. Candidate findings are validated by comparison, removing issues that reflect a project's general conventions rather than a unique violation, and finally the sink, the violated security condition, and the evidence are reported, making findings actionable and traceable.