← LLM Agent Security Vulnerabilities & Defense Research
Page History
Every revision of this knowledge page since it was created on September 16, 2026. Each entry shows what kind of change was made, how the page grew or shrank, and the source articles that drove it — so the paper's memory is inspectable, not taken on faith.
Updated from new reportingOctober 21, 2026 (+606 chars)
- SoK: PIAtlas: Decomposing the Unbounded Prompt Injection Attack Space
- Over-Defense in Coding Agents Begins at Silent Forks
- Cortex: A Fixed-Point Theory of Governed Coding Agents - Technical Report
- SPLASH: Contrastive Audit Pressure for Detecting and Localizing Security Hallucinations in Generated Code
- Many Repairs, One Direction: Steering Code Models Toward Safer Cryptographic API Use
Updated from new reportingOctober 12, 2026 (+1,366 chars)
Updated from new reportingSeptember 25, 2026 (+1,331 chars)
Updated from new reportingSeptember 23, 2026 (+901 chars)
Updated from new reportingSeptember 22, 2026 (+962 chars)
Updated from new reportingSeptember 20, 2026 (+977 chars)
Page createdSeptember 16, 2026 (+2,406 chars)
- ActGuard: Pre-execution Action Auditing against Indirect Prompt Injection in LLM Agents
- PIDS-Bench: Evaluating Prompt-Injection Detectors Under Over-Defense, Obfuscation, and Distribution Shift
- Pick Your Poison: Learning to Select Poison Sets for Stronger LLM Backdoor Attacks
- SpliTEE: Improving LLM Inference on Trusted Hardware with Differentially Private GPU Outsourcing
- SpliTEE: Improving LLM Inference on Trusted Hardware with Differentially Private GPU Outsourcing
- Mirror, Mirror on the Wall: Prompt Echoing in Small Instruct Language Models
- …and 4 more source articles