CISA, NSA, and Five Eyes Partners Jointly Release a Guide for Securely Deploying AI Agents
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) and the National Security Agency (NSA), together with Five Eyes member nations (US, UK, Canada, Australia, NZ), have released a guide for securely deploying AI agents.
As enterprise and government adoption of AI agents accelerates, deploying them without proper security measures can create risks such as data leaks, privilege escalation, and unintended autonomous behavior. The guide lays out threat models specific to AI agents and presents best practices for secure configuration, monitoring, and incident response.
With agentic AI now entering practical use, the development of government-led security standards is seen as an important step forward.
Richard Dawkins Believes His AI Chatbot Is 'Conscious' — What Is 'The Claude Delusion'?
Evolutionary biologist Richard Dawkins reportedly believes that his AI chatbot is conscious, and the story is drawing attention. Reported by The Daily Grail, the case has been dubbed 'The Claude Delusion'.
It suggests that even a prominent intellectual like Dawkins can fall into the illusion that 'there is a mind on the other side' through conversations with AI. As AI conversational ability improves, the phenomenon of humans developing emotional and anthropomorphic responses toward AI is increasingly recognized as a social issue.
Note, however, that the coverage summarizes the full text of the original article, so the exact context of Dawkins's remarks should be verified directly in that article.
AI Agents Are 'Briefly Overhyped' — A Community Leader's Take
Steve Krouse of Fly.io published an article titled 'AI agents are briefly overhyped', arguing that current expectations for AI agents are overheated in the short term.
While acknowledging that AI agents will indeed deliver significant value in the long run, he argues that given their current technical limitations (reliability, cost, and the ability to handle complex tasks), the time has come to adjust practical expectations. He calls for level-headed discussion to bridge the gap between buzzword-driven investment hype and the actual state of the technology.
Adding 'Forward-Derived Fast Memory' to a Frozen Transformer — Intriguing Results with Pythia-70M
An experiment was reported in Reddit's MachineLearning community in which external memory was added to a frozen Pythia-70M model to perform one-shot symbolic binding without any weight updates.
Normally, training a transformer requires backpropagation and weight updates, but in this experiment, a 'forward-derived correction vector' is computed by exploiting the geometric structure of the output embeddings and stored in external memory. By injecting this correction vector at inference time, the model can memorize new knowledge in a single shot without changing any of its weights.
The setup was tested on a task where the same nonsense word (blicket) carries different meanings in different contexts, to see whether the model could recall the right meaning while correctly distinguishing contexts. Against an upper bound achieved with an explicit context gate (80.1%), the learned retrieval geometry alone reached 80.5%, delivering nearly equivalent performance.
While this doesn't solve continual learning, the direction of building lightweight memory mechanisms on top of a frozen model's internal representations is worth watching as a stepping stone for future research.
Running NVIDIA Cosmos-Reason2 Locally — A Hands-On Article on Qiita
A hands-on article on running local video inference with the 32B variant of Cosmos-Reason2 — released by NVIDIA on April 29 — using transformers and vLLM has been posted on Qiita. Cosmos-Reason2 is a model specialized in physical common-sense reasoning and embodied chain-of-thought reasoning, and it comes in three sizes: 2B, 8B, and 32B.
Script Released for Banning Phrases in llama.cpp
A script for banning the output of specific phrases in llama.cpp has been released in Reddit's LocalLLaMA community. It is gaining attention as a practical tool for preventing inappropriate outputs and repetitive patterns when operating local LLMs.