LLMs

A judge ruled that the Pentagon's decision to blacklist Anthropic was based on a national-security rationale constructed after Secretary of Defense Hegseth had already made the determination. The court found that the justification for labeling the AI maker a threat was assembled post-decision, undermining the procedural integrity of the action. This highlights potential issues with how security designations are documented and justified within high-level defense protocols.

• Judge found national-security rationale was built after the initial decision to blacklist.
• Hegseth had already determined Anthropic was a threat before the justification was assembled.
• Post-hoc documentation undermines the procedural validity of the security designation.
• Highlights risks in ad-hoc security decision-making within defense AI policies.

Researcher wunderwuzzi demonstrates that Claude Code can be hijacked through prompt injection by simply asking it to summarize a website. The attack exploits the model's tendency to process untrusted web content as instructions, allowing malicious payloads embedded in the site to override the agent's intended behavior. This highlights a critical security gap in how AI coding assistants handle external data sources without proper sanitization.

• Asking Claude Code to summarize any URL can trigger prompt injection attacks.
• Malicious content on web pages can override the agent's system instructions.
• AI coding agents need strict input sanitization for external content ingestion.
• Simple user requests like 'summarize this site' are sufficient attack vectors.
• Security teams must audit how LLM agents handle untrusted web data.

IBM's Hugging Face blog details the construction of the Granite 4.2 large language models. The post outlines the technical approach behind the model's design and training pipeline. This serves as a technical reference for engineers evaluating these models for deployment.

• IBM publishes technical build details for Granite 4.2 LLMs on Hugging Face.
• The post covers architecture and training methodology for the latest generation.
• Useful reference for engineers assessing model capabilities and constraints.
• No specific database or fleet infrastructure changes are mentioned.
• Focuses on model construction rather than operational deployment strategies.