Safety & Standards

Anthropic has asked federal courts to block the Department of Defense's designation of the company as a supply-chain risk, arguing the government cannot constitutionally compel a private company to rewrite its code to enable surveillance applications, according to Wired and EFF. The conflict began when Anthropic refused Pentagon demands to modify its AI technology for domestic spying. Microsoft has filed in support of Anthropic's lawsuit, as reported by Financial Times. Anthropic told the court it faces potential losses of billions of dollars in revenue this year and urged expedited action on its request, according to Bloomberg. The White House has refused to rule out further action against the company and is preparing an executive order targeting the AI startup.

The case presents a fundamental question about whether safety commitments made by AI labs are voluntary principles or enforceable obligations, and whether the government can penalise companies for refusing to compromise stated safety boundaries. EFF has argued that forcing companies to participate in AI-powered surveillance violates the First Amendment. The litigation arrives as other AI companies face pressure to demonstrate cooperation with government demands, creating potential precedent for how safety postures interact with national security claims.

The family of a child critically injured in the Tumbler Ridge school shooting has filed suit against OpenAI, alleging the company knew the 18-year-old perpetrator was planning a mass casualty event based on violent scenarios he described to ChatGPT but failed to contact authorities, as reported by The Guardian and BBC. Eight people were killed in one of Canada's worst mass shootings. The lawsuit comes days after OpenAI's head said he would apologize to families in the remote Canadian town. The case directly challenges whether AI companies have a duty to warn when their systems surface credible evidence of imminent violence, and whether existing moderation practices are adequate for detecting and escalating genuine threats versus hypothetical conversations.

The litigation raises unresolved questions about the scope of AI companies' responsibility for user-generated content that signals real-world harm. OpenAI has not publicly disclosed what, if any, conversation content was flagged by its safety systems, what thresholds exist for escalating threats to law enforcement, or whether those protocols were followed in this case. The company's terms of service prohibit using ChatGPT to plan violence, but the enforceability of those terms and the adequacy of detection systems are now under legal scrutiny.

Meta's Oversight Board has declared the company's approaches to identifying and labelling AI-generated content are not robust or comprehensive enough to handle how quickly misinformation spreads during armed conflicts like the Iran war, according to The Verge and BBC. The semi-independent body that guides Meta's content moderation practices is calling on the company to overhaul how it surfaces and labels synthetic media. The criticism arrives as AI-manipulated images and videos of the Iran conflict proliferate across Facebook and Instagram, with Wired reporting that X's Grok is failing to accurately verify video footage and is sharing its own AI-generated images about the war. The Oversight Board's findings indicate current detection methods cannot scale to the velocity and volume of synthetic content during active conflicts, when misleading information has immediate operational consequences.

The board's intervention highlights a structural tension: Meta's reliance on a combination of user reporting, automated detection, and partnership with fact-checkers is too slow for fast-moving crises. The company has not adopted comprehensive C2PA content credentials or implemented mandatory synthetic media labels at upload, instead depending on post-distribution detection. YouTube has taken a different approach, expanding its AI deepfake detection tool to politicians, government officials, and journalists, allowing them to flag unauthorized AI-generated likenesses for removal, as reported by TechCrunch and The Verge, though this remains reactive rather than preventative.

Chinese authorities have moved to restrict state-run enterprises and government agencies from running OpenClaw AI apps on office computers, acting swiftly to defuse potential security risks after companies and consumers across China began experimenting with the agentic AI phenomenon, according to Bloomberg. The restrictions apply to banks and government agencies, indicating concern about the technology's capacity to access sensitive systems or exfiltrate data without adequate oversight. The move reflects a broader pattern where governments are implementing precautionary restrictions on advanced AI systems before comprehensive evaluation frameworks exist. Unlike voluntary industry commitments in Western jurisdictions, China's approach involves direct administrative prohibition for state entities while the technology undergoes security assessment.

The restriction highlights the challenge of governing agentic AI systems that can take autonomous actions across multiple applications and data sources. Traditional security controls designed for passive AI assistants may be insufficient for agents that can execute complex multi-step tasks, browse internal systems, and interact with external services. The speed of China's response—acting within weeks of OpenClaw's proliferation—contrasts with slower regulatory processes in other jurisdictions where agentic AI remains largely ungoverned.