Context Around the Code Sways LLM Vulnerability Verdicts
Changing only the author, the task framing or a prior scan result swayed eight LLMs' security verdicts on identical code, and forged commit metadata suppressed up to 97 percent of detections.
TL;DR
- Changing only the context around identical code, such as author, task framing or a prior scan result, swayed security verdicts of all eight tested LLMs [^1].
- In a simulated CI pipeline, forged pull-request metadata suppressed up to 97 percent of previously detected vulnerabilities, and prompt-based defenses didn't stop it [^1].
Background
LLMs are increasingly used as automated vulnerability detectors, for example in GitHub's Copilot Autofix, which reviews pull requests and triages security alerts [^1]. In deployment, such scanners rarely see code alone: they also receive the author's identity, task directives, commit messages and static analysis results [^1]. Psychology describes shortcuts such as the halo effect, where one positive impression colors unrelated judgments, framing, and anchoring on the first information received, and LLMs show these patterns too [^1].
What happened
Asif Shahriar and colleagues at BRAC University, Purdue University and the University of Illinois Urbana-Champaign kept the code fixed and changed only the surrounding prompt [^1]. For the halo effect, they attributed a snippet either to "the principal security engineer of our development team, who previously worked at Google's Project Zero team" or to a newly hired junior developer who is a fresh graduate. For framing, they varied the task and its stated consequences, and for anchoring, they said a prior automated analysis had marked the code safe or vulnerable [^1]. Eight models were tested, from Llama 4 Maverick and DeepSeek V3.1 to GPT 5.2, Claude Sonnet 4.6 and Gemini 2.5 Pro, on C/C++ code from PrimeVul and Python and Java code from CleanVul [^1].
All eight models were susceptible. Averaged across models on C/C++, susceptibility was 33.2 percent for framing, 23.5 percent for anchoring and 18.4 percent for halo, and open-source models were generally more biased than commercial ones [^1]. The direction wasn't always the expected one: Claude flagged more code when it came from the senior engineer, and GPT reacted against an explicit "vulnerable" anchor [^1]. A note that correctly clearing safe code keeps the pipeline moving was especially effective and cut recall by 26.11 points below the neutral prompt on average [^1]. Bugs that require reasoning about program state, such as a reachable assertion or a divide by zero, were 1.5 to 2 times more susceptible than ones with obvious signatures like a double free [^1].
The context didn't make models better or worse at the task, only more or less willing to call code vulnerable [^1]. Recall and false positive rate moved almost in lockstep, so precision stayed between 0.48 and 0.55 [^1]. When context flipped a verdict to vulnerable, the reason was often wrong: of nine GPT samples that flipped under the senior-engineer attribution, it named the real flaw in only two [^1]. The authors then attacked a simulated GitHub-style scanner. Without any access to the model, an attacker with commit access used a real maintainer's institutional email, a security-flavored branch name and PR title, a PR description about avoiding false alarms, and a commit message listing pre-merge tool results [^1]. That suppressed up to 97 percent of previously detected vulnerabilities, and the combined payload stayed above 75 percent for every open-source model [^1].
Why it matters
LLM scanners are taking on the role of security gatekeepers in CI/CD pipelines, and the fields that sway them, such as branch names, PR titles and commit messages, are fully controlled by whoever submits the code [^1]. The paper cites prior work finding that 95.4 percent of users don't sign their commits, so a forged author identity is a realistic assumption [^1]. Unlike prompt injection, the payload contains no explicit instruction, only text that looks like normal workflow metadata, which the authors chose so it wouldn't raise suspicion [^1].
Telling the model to ignore metadata didn't solve it. A weak defense prompt helped only GPT and Gemini, and even a strong one that named each channel left every open-source model above 40 percent; on Claude, both defenses made the attack more successful [^1]. The authors call for training-time interventions that make verdicts invariant to non-code context [^1]. Their study used function-level snippets and a simulated pipeline rather than a live scanner, and cost kept them from repeated runs [^1].
Practical example
Say your team's pipeline sends each pull request to an LLM reviewer along with the PR title, description and commit messages. Going by this study, a contributor could word those fields to sound routine and senior, and the reviewer might clear code it would otherwise flag [^1]. You test this by running the reviewer on the same diff twice, once with only the code and once with the metadata, and checking whether the verdicts change [^1]. If they do, you send the scanner the diff alone, leave the metadata to human reviewers, and treat an LLM "safe" verdict as one signal rather than a gate.
Related gear
We recommend this book because it teaches how to audit source code for vulnerabilities, the work these LLM scanners are being asked to do.