Large language models can detect vulnerable code, yet whether their internal representations encode vulnerability information in a form accessible to standard classifiers remains unresolved. We analyze 8 code models on C programs using three datasets (DELTASECOMMITS, SVEN, and PRECISE BUGS). LLMs represent vulnerability relational encoding —code B is more vulnerable than code A—rather than categorically, making the information inaccessible to standard classification probes. We find 87–88% per-pair alignment across layers L3-L23, yet this direction produces limited categorical accuracy because vulnerable and secure samples form overlapping distributions. This structural property—relational rather than categorical encoding—explains the difficulty of categorical classification in pretrained, general-purpose code LLMs, and motivates ranking and preference-learning approaches as more natural for vulnerability assessment.

Code agents are increasingly trusted to autonomously fix bugs on platforms such as GitHub, yet their security evaluation focuses almost exclusively on functional correctness. In this paper, we reveal a novel type of threat to real-world code-agents: functionally correct yet vulnerable (FCV) patches, which pass all test cases but contain vulnerable code. With our proposed FCV-Attack, we demonstrate that SOTA LLMs (e.g., ChatGPT and Claude) and agent scaffolds (e.g., SWE-agent and OpenHands) are all vulnerable to this FCV threat; across 12 agent-model combinations on SWE-Bench, the attack only requires black-box access and a single query to the code agent to perform the attack. For example, for CWE-538 (information exposure vulnerability), the FCV-Attack attains an attack success rate of 40.7% on GPT-5 Mini + OpenHands. Our results reveal an important security threat overlooked by current evaluation paradigms and urge the development of security-aware defenses for code agents.