Prompt-Injection Cascades in a Multi-Agent Enterprise (ATLAS)
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.
- from
- Prompt Injection
- added
- 2026-10-10
- likes
- 0
similar
-
The Landscape of Prompt Injection Threats in LLM Agents (SoK) arxiv.org
Feb 2026 systematization-of-knowledge paper with a unified taxonomy covering attack payload strategies (heuristic vs. optimisation-based) and defense intervention stages (text, model, execution). Introduces the AgentPI benchmark for context-dependent agent tasks that all prior…
-
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…
-
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.
-
Continuously Hardening ChatGPT Atlas Against Prompt Injection Attacks openai.com
OpenAI's Dec 2025 disclosure of a real attack chain (malicious email → agent sends resignation letter) and the RL-trained automated attacker they built to find new injection classes before external adversaries do. OpenAI explicitly states deterministic guarantees are not…
-
New Prompt Injection Papers: Agents Rule of Two and The Attacker Moves Second simonwillison.net
Simon Willison's Nov 2025 commentary on both landmark papers, including the finding that 12 published defenses were bypassed at >90% success rate using gradient descent and RL-based adaptive attacks.
-
Prompt Injection 2.0: Hybrid AI Threats arxiv.org
Jul 2025 paper showing how prompt injections now combine with XSS, CSRF, AI worm propagation, and multi-agent infections to evade traditional WAFs entirely. Evaluates Preamble's classifier, data-tagging, and RL-based defenses against these hybrid scenarios.
Prompt Injection › Articles and Blog posts: “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.”