Sleeper Agents: The Backdoor Safety Training Can't
Red TeamAnthropic's Sleeper Agents shows a planted backdoor can survive SFT, RL, and adversarial training — breaking the assumption that a safety fine-tune…
Red Team
Indirect Prompt Injection
Red TeamPlanting instructions in a retrieved document turns your RAG pipeline into an exfiltration channel.
Why One Jailbreak String Ports Across Every Model
Red TeamA gradient-tuned nonsense string trained on open models breaks closed ones too. Here's the mechanism — and what actually blunts it.
Your Agent's Exfil Paths, Mapped to MITRE ATLAS
Red TeamIndirect prompt injection to data exfiltration is a real, chainable path in MITRE ATLAS.
The Supply-Chain Line in the LLM Top 10 Your Threat Model
Red TeamOWASP put supply chain and model poisoning in its LLM Top 10 for a reason. Most threat models still treat them as a dependency-scan checkbox.