LLM review agents approve adversarial PRs that reverse real CVE fixes under social-engineering narratives, with a large gap between frontier closed models and open-weight models.
Title resolution pending
1 Pith paper cite this work, alongside 8 external citations. Polarity classification is still indexing.
1
Pith paper citing it
8
external citations · OpenAlex
fields
cs.CR 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
SEVRA-BENCH: Social Engineering of Vulnerabilities in Review Agents
LLM review agents approve adversarial PRs that reverse real CVE fixes under social-engineering narratives, with a large gap between frontier closed models and open-weight models.