The research presented in this CAMLIS talk builds a guardrail from a frozen language model's hidden activations instead of fine-tuning a classifier. It caches representations of labelled safe and unsafe examples, compares incoming prompts with nearby examples across layers, and combines that signal with sentence embeddings. The experimental bank contains 50 examples per class, not 50 total. Reference-bank and evaluation datasets are separated across six topical and security domains. Updating the bank changes the classification task without changing model weights, but still requires labelled examples, activation access and validation. Domain-specific system prompts change the balance between false alarms and missed attacks. The paper's latency comparisons use a particular GPU and model setup, and stronger average scores do not establish resistance to every adaptive attack. The operational question is whether a chosen bank, prompt and threshold provide acceptable errors on the traffic the guardrail will actually see.
kNNGuard: evaluate activation-based guardrails with separate reference and test data
Related research
More curated notes connected through Prompt Injection and Model Evaluation.
Pacing model development in an era of cyber-critical capabilities
OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.
OWASP Top 10 for Agentic Applications for 2026
OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.
A blueprint for democratic governance of frontier AI
OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.