…prompt injection. We use techniques like spotlighting that direct the model to strongly prefer following user and system instructions over what’s on the page, and we’ve upstreamed known attacks to…
…Mandiant in late 2025 responded to an attack on a KnowledgeDeliver server and says that initially, the vulnerability was exploited as a zero-day to inject a malicious script into the web…
…Adversarial training could be identifying prompts that are used to trick you and then training your model to recognize them. However, attackers are always coming up with new tricks, so prompt injection…
…The current state of prompt injections on the web We initiated a broad sweep of the public web to monitor for known indirect prompt injection patterns. This is what we found. By…
Interesting new research you may have heard of on attacking large audio language models. The attack is called AudioHijack and the part worth paying attention to is that adversarial clips built against open models transfe…
Hey HN! We're Dr. Kashyap Thimmaraju and Giuseppe Canale from Silicon Psyche. We've built Posture Sequence Analysis (PSA), a behavioural health monitor for LLMs and AI Agents.Why we built PSAWe built PSA because we wante…
…Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced…
…Prompt injection . Instructions planted in a file, webpage, or tool output hijack the agent, redirecting it from the user's task toward the attacker's. A misaligned model . Canonically, misalignment occurs when…
…Seeing as it was only two months ago that Prompt Armor warned attackers could exfiltrate files from Claude Cowork via indirect prompt injection, it might be wise to take Microsoft's reassurances…
…a talented security researcher — @xiridium! Latest Safeguarding VS Code against prompt injections When a chat conversation is poisoned by indirect prompt injection, it can result in the exposure of GitHub tokens, confidential…
…Over half of AI prompt inputs are sent to personal accounts, and 23% of sensitive prompt uploads involve data transiting through personal or unverified accounts (i.e., outside the reach of any…
…Reinforcement Learning-based Red Teaming for Prompt Injection Defenses (2026) Uncovering Linguistic Fragility in Vision-Language-Action Models via Diversity-Aware Red Teaming (2026) T-MAP: Red-Teaming LLM Agents with Trajectory…