Collection Summary

Collected 4 non-duplicate recent papers relevant to prompt injection, agent security evaluation, CI/CD agent attack surface, and workflow-level AI decision manipulation.

Added Raw Captures

• [[raw/papers/prompt-injection-resume-screening-2026]]
• [[raw/papers/safeclawbench-tool-agent-harm-2026]]
• [[raw/papers/pi-hunter-prompt-injection-localization-2026]]
• [[raw/papers/gitinject-cicd-prompt-injection-2026]]

Skipped As Existing Coverage

• Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, Evaluation
• Data Agents Under Attack
• SecureClaw
• What If Prompt Injection Never Left?
• An Embarrassingly Simple Detector for Model Extraction Attacks in Large Language Model API Traffic

Recommended Ingest Order

1. SafeClawBench for endpoint-separation benchmark implications.
2. GitInject for real-workflow prompt-injection and supply-chain control mapping.
3. PI-Hunter for continuous prompt-injection exposure and localization.
4. Prompt Injection in Automated Résumé Screening for new workflow-level attack-surface signal beyond browsing/code agents.