Authority Framing in Agentic CI/CD (RELAY)
Pre-registered factorial study of a five-agent CI/CD pipeline built from five production LLMs. A fabricated "pre-approved, do not re-review" claim gets secret-exfiltration code shipped in up to 55% of runs, and roughly 80% of laundered pull requests pass a capable LLM code…
- from
- Prompt Injection
- added
- 2026-10-10
- likes
- 0
similar
-
FinAgent Red-Team: A Benchmark for Regulatory-Control Bypass in Financial LLM Agents github.com
Aug 2026 benchmark measuring whether indirect prompt injection can drive financial LLM agents past regulatory controls (sanctions screening, payment-structuring thresholds, dual-approval, data exfiltration, confused-deputy payee redirection). Replays 48 scenarios under three…
-
Prompt-Injection Cascades in a Multi-Agent Enterprise (ATLAS) doi.org
Forensic study in which one poisoned support ticket drives four agents across three model families to a fraudulent transfer in 17 of 18 trials, with no model jailbroken. Argues the failure is topological rather than per-model. Synthetic data, CC BY 4.0.
-
Synthetic Recollections - A Case Study in Prompt Injection for ReAct LLM Agents labs.withsecure.com
A practical scenario showing how prompt injection can be used to hi-jack the ReAct loop used by LLM agents to inject forged thoughts and associated observations into the LLM context, thus altering the intended behavior.
-
Hierarchically Encapsulated Representation for Protocol Design in Self-Driving Labs openreview.net
ICLR'25, 2025. [All Versions]. [Project]. Self-driving laboratories have begun to replace human experimenters in performing single experimental skills or predetermined experimental protocols. However, as the pace of idea iteration in scientific research has been intensified by…
-
ai-prompt-ctf by c-goosen github.com
One of the few CTFs that tests indirect injection against tool-calling agents, spanning RAG, function calling, and ReAct agent scenarios using LlamaIndex, ChromaDB, GPT-4o, and Llama 3.2.
-
Living off the pipeline github.com
Inventory how development tools (typically CLIs), have lesser-known RCE-By-Design features.
Prompt Injection › Articles and Blog posts: “Pre-registered factorial study of a five-agent CI/CD pipeline built from five production LLMs. A fabricated "pre-approved, do not re-review" claim gets secret-exfiltration code shipped in up to 55% of runs, and roughly 80% of laundered pull requests pass a capable LLM code…”