OpenAI has acquired Promptfoo, a startup specializing in prompt engineering and evaluation tools, to enhance the security and reliability of its AI agents. This acquisition underscores the growing necessity to ensure the safety and predictability of AI agents as they undertake increasingly complex tasks.

Promptfoo's technology facilitates the testing, evaluation, and comparison of prompts, aiding developers in systematically improving the quality and safety of AI model outputs. Through this integration, OpenAI aims to establish a robust framework to ensure its LLM-powered agents perform as intended, minimizing undesirable behaviors and security risks.

👉 Read the full article on TechCrunch

OpenAI details how it leverages "skills" in conjunction with Codex and GitHub Actions to accelerate maintenance of its Agents SDK repositories. This approach transforms repetitive engineering tasks such as verification, release preparation, integration testing, and PR review into repeatable workflows, significantly boosting development throughput.

A "skill" encapsulates repository-specific operational knowledge, comprising a SKILL.md manifest, optional scripts, references, and assets. This provides Codex with stable context about how the repository functions, improving the speed and accuracy of recurring engineering work. Notably, by setting rules in the AGENTS.md file to mandate specific skill usage, the reliability of workflows is enhanced.

The article showcases concrete skills and their implementation patterns, including code change verification, documentation synchronization, automated example runs, release reviews, and PR draft preparation. It also emphasizes the importance of detailed description fields for skills and the best practice of delegating deterministic shell work to scripts while allowing the model to focus on interpretation, comparison, and reporting.

👉 Read the full article on OpenAI Developers

Anthropic has filed two lawsuits against the U.S. Department of Defense for labeling the company a "supply chain risk to national security." This designation effectively prohibits Anthropic from securing U.S. government contracts and blacklists it among other defense contractors.

The Pentagon's decision stemmed from Anthropic's refusal to agree to new terms that would permit the U.S. government to use its AI model, Claude, for mass domestic surveillance and the development of fully autonomous weapons. Anthropic argues that these actions are "unprecedented and unlawful," asserting that the Constitution does not grant the government the power to punish a company for its "protected speech."

Anthropic clarifies that Claude has never been tested for such uses, and its guardrails are rooted in the company's understanding of the model's risks and limitations. The lawsuits are expected to have significant implications for the broader debate on AI safety and military applications, with concerns raised about potential impacts on U.S. AI competitiveness.

👉 Read the full article on Gizmodo