Safety & Standards
Top Line
The Pentagon formally designated Anthropic a supply-chain risk—the first such label for a U.S. company—after weeks of failed negotiations over acceptable use policies for military AI deployments, with Anthropic vowing legal action while reportedly attempting to resume talks.
OpenAI released GPT-5.4 with native computer control capabilities and announced Pentagon deals to fill the gap left by Anthropic, triggering a 300% spike in ChatGPT uninstalls from users and employees opposed to supporting government surveillance programmes.
The U.S. is reportedly drafting export control rules that would require countries to invest in America in exchange for advanced AI chips from Nvidia and AMD, representing sweeping new government authority over global semiconductor sales regardless of origin country.
Meta faces lawsuits after investigations revealed that contractors in Kenya reviewed sensitive footage from AI smart glasses—including nudity and sex—contradicting marketing promises of privacy and user control over data sharing.
North Korean IT workers are using AI-powered voice-changing tools and identity masking to secure remote positions at Western firms, with wages funnelled back to Pyongyang in an evolved signature money-raising scheme, according to Microsoft.
Key Developments
Pentagon-Anthropic Standoff Escalates to Legal Battle and Supply-Chain Designation
The Department of Defense formally notified Anthropic that it has designated the company and its products as a supply-chain risk, according to Bloomberg and The Verge, marking the first time this label—historically reserved for foreign adversaries like Huawei—has been applied to an American company. Anthropic CEO Dario Amodei stated the company has no choice but to challenge the designation in court, calling it unprecedented and claiming the vast majority of Anthropic's customers remain unaffected, per Financial Times. Multiple sources indicate talks between the parties have restarted despite the formal designation, according to TechCrunch, though the breakdown centres on Anthropic's refusal to grant unrestricted military access to its Claude AI models, particularly for surveillance and autonomous weapons systems.
The dispute has drawn criticism from tech lobbyists, investors, and former Trump advisers who warn the administration is undermining its own deregulatory, export-driven AI agenda by hammering a top U.S. company, as reported by Politico. Meanwhile, the Pentagon continues using Anthropic's AI in operations against Iran even as it labels the company a risk, creating an operational contradiction. The standoff has also drawn attention to the lightly regulated practice of government purchase and AI-enhanced analysis of commercially available personal information including browsing histories and location data, according to Bloomberg.
OpenAI Fills Military Gap With Pentagon Deal and Computer Control Capabilities
OpenAI launched GPT-5.4, its first model with native computer use capabilities allowing it to operate across applications autonomously, alongside new financial services tools designed for professional work, as reported by TechCrunch and The Verge. The release coincided with announcements of Pentagon deals to provide AI capabilities following Anthropic's refusal, with Bloomberg noting OpenAI is positioning itself as the military's primary AI provider. However, sources told Wired that the Defense Department had already been testing OpenAI models through Microsoft before the company officially lifted its military use prohibition.
The military pivot triggered immediate backlash, with early reports showing ChatGPT uninstalls rose nearly 300 percent after the Pentagon deals were announced, according to Bloomberg, with both users and employees objecting to supporting government mass surveillance. Downloads of rival Anthropic's Claude surged during the same period. The Electronic Frontier Foundation argues that OpenAI's commitments contain weasel words that fail to prevent AI-powered surveillance, noting the company's acceptable use policies provide insufficient safeguards against mass data collection and analysis by government agencies.
Meta Faces Legal Action Over Smart Glasses Privacy Violations
Meta is facing lawsuits after investigations by Swedish outlets Svenska Dagbladet and Göteborgs-Posten revealed that contractors in Nairobi, Kenya reviewed sensitive footage from customers' AI-powered smart glasses, including videos showing bathroom visits, sex, and other intimate moments, according to TechCrunch and The Verge. The footage review directly contradicts Meta's marketing materials that promised privacy and user control over when footage would be shared, according to lawyers bringing the complaints. The glasses capture continuous video that users believe remains private unless explicitly shared, but the investigation found systematic human review of content for training and quality assurance purposes.
The case highlights a fundamental gap between AI safety marketing and actual data handling practices. Meta has not disclosed the volume of footage reviewed, the criteria for selecting videos for human analysis, or the security protocols protecting sensitive content during the review process. The Kenya location for content review raises additional questions about labour practices and data protection compliance, particularly given the EU's GDPR and similar privacy frameworks that govern European users whose footage may be included in the review pools.
U.S. Drafts Sweeping Export Control Framework for AI Chips
The U.S. government is considering draft rules that would require countries to invest in America in exchange for access to advanced AI chips from Nvidia and AMD, representing unprecedented government authority over global semiconductor sales, according to Financial Times and Bloomberg. The proposed framework would give the U.S. a formal role in approving chip exports regardless of which country manufactures them, tying access to reciprocal investment commitments. Ars Technica notes the framework includes pledges from data centre companies to fund their own power generation, though those commitments lack enforcement mechanisms and face questionable economics.
The proposal represents a sharp expansion of U.S. extraterritorial authority, effectively positioning Washington as a gatekeeper for global AI compute access. The draft rules come as the administration simultaneously pursues aggressive AI export promotion and technology leadership, creating tension between open markets and strategic control. Industry sources tell TechCrunch that the permit requirement would apply to every chip export sale, creating massive new compliance burdens and centralising control in a way that may prove administratively unworkable.
Signals & Trends
Safety Restrictions Function as Negotiable Commercial Terms, Not Technical or Ethical Constraints
The Pentagon-Anthropic standoff and OpenAI's military pivot reveal that AI safety commitments operate as negotiable business terms rather than binding technical or ethical boundaries. Companies can remove restrictions when commercial or political pressure becomes sufficient, while those maintaining restrictions face designation as national security risks. This pattern suggests current voluntary frameworks provide no meaningful constraint on deployment in high-stakes government applications. The market is learning that safety positioning functions primarily as a differentiation strategy until government demands override it, at which point compliance becomes mandatory regardless of stated principles. This creates adverse selection where companies with strongest safety cultures face punishment while those willing to accommodate military demands without restriction gain competitive advantage.
Privacy Marketing Claims Face Growing Legal Liability as Actual Practices Are Exposed
The Meta smart glasses lawsuit joins a pattern of consumer AI products whose actual data handling contradicts privacy marketing. Similar cases exist around AI agents accessing emails, voice assistants recording conversations, and AI image tools retaining training data. The gap between marketing promises and operational reality is not accidental but structural—AI systems require data labelling, quality control, and failure analysis that necessitates human review, yet disclosing this undermines consumer trust. As investigations expose these practices, companies face legal liability both for misleading marketing and for privacy violations. The trend suggests privacy promises for AI products should be treated with extreme scepticism by both consumers and compliance teams until independent audits verify actual data flows match stated policies.
AI Safety Governance Increasingly Functions Through Procurement Power Rather Than Regulation
The Pentagon's ability to coerce Anthropic through procurement decisions and supply-chain designations—while circumventing formal rulemaking—demonstrates how government purchasing power creates de facto regulation without democratic oversight or due process. Similar patterns appear in government use of commercially available data, law enforcement access to AI tools, and intelligence agencies' AI procurement. This procurement-based governance operates faster than legislation, faces minimal judicial review, and concentrates authority in executive agencies. It creates asymmetric power where government can impose requirements on companies through contract terms and access denial that would face legal challenge if implemented as regulation. The trend suggests meaningful AI governance debate must focus on procurement rules and government buyer power, not just formal regulation.
Explore Other Categories
Read detailed analysis in other strategic domains