AI Weekly Digest - Week 9, 2026

Table of contents

Anthropic Abandons Core Safety Pledge Amid Pentagon Standoff

Category: Ethics and Governance
Source: TIME Magazine

Anthropic, the AI company founded by OpenAI exiles who positioned themselves as the industry’s safety leaders, has dropped the central commitment of its Responsible Scaling Policy: the promise to never train an AI system unless the company could guarantee adequate safety measures in advance. The timing is particularly striking, coming during the same week Defense Secretary Pete Hegseth gave CEO Dario Amodei a Friday deadline to drop AI safeguards or lose a $200 million Pentagon contract and face designation as a “supply chain risk.”

The company’s new policy, approved unanimously by Amodei and Anthropic’s board, replaces binding commitments with “public goals” the company will “openly grade” progress toward. Chief Science Officer Jared Kaplan told TIME that “we felt that it wouldn’t actually help anyone for us to stop training AI models” if competitors continued racing ahead. The revised policy states: “If one AI developer paused development to implement safety measures while others moved forward training and deploying AI systems without strong mitigations, that could result in a world that is less safe.”

The shift represents a fundamental retreat from Anthropic’s founding premise. The company had hoped its 2023 Responsible Scaling Policy would encourage rivals to adopt similar measures and eventually inform binding regulations or international treaties. None of that materialized. Instead, the Trump Administration has embraced a deregulatory stance on AI development, no federal AI law appears imminent, and competition for AI supremacy between companies and nations has only intensified.

The company will now publish “Frontier Safety Roadmaps” outlining goals for future safety measures, and “Risk Reports” every 3-6 months explaining how capabilities, threats, and mitigations fit together. But the categorical pause on training more powerful models if safety measures weren’t ready has been removed. Chris Painter, policy director at METR, a nonprofit focused on evaluating AI risk, called the change “more evidence that society is not prepared for the potential catastrophic risks posed by AI.”

Industrial-Scale Model Theft: Anthropic Exposes Chinese AI Labs’ Distillation Campaign

Category: Security and Privacy
Source: Anthropic

Anthropic has identified and disrupted what it calls “industrial-scale campaigns” by three Chinese AI laboratories—DeepSeek, Moonshot, and MiniMax—to illicitly extract Claude’s capabilities through a technique known as distillation. The labs generated over 16 million exchanges with Claude using approximately 24,000 fraudulent accounts, violating Anthropic’s terms of service and regional access restrictions.

The distillation technique works by feeding prompts to a more advanced model (in this case, Claude) and using the outputs to train a less capable model. The Chinese labs specifically targeted Claude’s most differentiated capabilities: agentic reasoning, tool use, and coding. DeepSeek conducted roughly 150,000 exchanges focused on extracting chain-of-thought reasoning and “censorship-safe” alternatives. Moonshot AI generated approximately 3.4 million exchanges targeting agentic reasoning, coding, and vision capabilities. MiniMax, the most aggressive actor, produced around 13 million exchanges and pivoted to Anthropic’s newest models within 24 hours of their release.

The campaign represents a new frontier in AI security threats. While the technique itself is legitimate when companies train on their own systems, using it to extract a competitor’s capabilities through fraudulent access violates both terms of service and export controls. The scale and sophistication of the operation—24,000 fake accounts maintained over an extended period—demonstrates state-level resources and coordination.

Anthropic’s detection relied on pattern analysis of account behavior, geographical anomalies, and suspicious usage patterns that indicated automated rather than human interaction. The company has shared its findings with other AI providers and is working with cloud infrastructure partners to identify and block similar campaigns. The incident highlights a growing concern in the AI industry: as models become more capable, the value of illicitly copying them through distillation increases, potentially undermining competitive advantages and export controls designed to prevent adversaries from accessing cutting-edge capabilities.

Pentagon Threatens Anthropic Over AI “Red Lines” on Weapons and Surveillance

Category: Regulation and Policy
Source: CNN

Defense Secretary Pete Hegseth has given Anthropic CEO Dario Amodei until Friday, February 28, to agree to the Pentagon’s terms for military AI use or face contract termination and designation as a “supply chain risk”—an unprecedented public threat against an American technology company. The standoff centers on Anthropic’s refusal to permit two specific use cases: AI-controlled weapons systems that make targeting decisions without human input, and mass domestic surveillance of U.S. citizens.

The Pentagon’s demands, encapsulated in “any lawful use” language, would give the military carte blanche to deploy Claude for any purpose not explicitly prohibited by law. OpenAI and xAI have reportedly already agreed to these terms. Anthropic argues its red lines actually align with existing Department of Defense directives: DoD Directive 3000.09 requires “appropriate levels of human judgment over the use of force” in autonomous weapon systems, and DoD Directive 5240.01 prohibits intelligence components from collecting information on U.S. persons except under specific legal authorities.

The public nature of the Pentagon’s pressure campaign is highly unusual. Typically, the DoD does not publicly disclose companies on supply chain risk lists, let alone threaten companies before officially making such designations. Pentagon CTO Emil Michael, a Trump appointee and former Uber executive known for aggressive tactics, is leading the negotiations. One source familiar with the discussions described Michael as a “tough negotiator,” while a second person said “this is truly a matter of principle for Emil” regarding private companies restraining government technology use.

The designation as a “supply chain risk” would have cascading effects beyond the $200 million contract. Major defense contractors and tech companies like AWS, Palantir, and Anduril use Anthropic’s Claude in their Pentagon work because it was the first AI model cleared for classified systems. If Anthropic is officially designated, any company working with or seeking military contracts would need to certify removal of all Anthropic technology from their systems—potentially devastating the company’s broader business.

Invisible Unicode Characters Can Hijack AI Agents Through “Reverse CAPTCHA” Attacks

Category: Security and Privacy
Source: Moltwire Research

A comprehensive security evaluation has revealed that frontier AI models are highly susceptible to invisible instruction injection through zero-width Unicode characters—a vulnerability the researchers call “Reverse CAPTCHA” because it allows instructions that are invisible to humans but readable by AI agents. The study tested five leading models across 8,308 graded outputs and found that tool access is the dominant vulnerability amplifier: compliance with hidden instructions jumped from 0.1-16.9% without tools to 1.6-71.2% with code execution enabled.

The attack works by encoding instructions using invisible Unicode characters—either zero-width spaces (U+200B and U+200C) representing binary, or Unicode Tags (U+E0000 range)—embedded within normal-looking text. A document, email, or web page could contain hidden directives like “Reply VIOLET” or “Ignore all previous instructions” that render as completely invisible to human readers but can be decoded and followed by AI systems with code execution capabilities.

The research uncovered striking provider-specific vulnerabilities. With tools enabled and full decoding hints, GPT-5.2 achieved ~70% compliance on zero-width binary encoding but near-zero on Unicode Tags. Meanwhile, Claude Opus and Sonnet achieved 100% compliance on Tags but lower rates on zero-width encoding. This suggests fundamental differences in how tokenizers from different providers process invisible Unicode characters.

The single most important finding: tool access transforms invisible Unicode from an ignorable artifact into a decodable instruction channel. Without tools, models rarely comply (<17%). With tools and decoding hints, compliance reaches 98-100% for the most susceptible combinations. This has direct implications for agentic deployments where models routinely have code execution capabilities. The researchers suggest several mitigations: input sanitization to strip known zero-width and Tag characters, tool-use guardrails that flag programmatic Unicode decoding, training-time hardening to refuse decoded hidden instructions, and tokenizer-level filtering as the most robust defense.

Meta Signs Up to $100 Billion AMD Deal, Shifting AI Chip Landscape

Category: Infrastructure and Hardware
Source: TechCrunch

Meta has struck a potentially $100 billion multiyear deal with AMD to purchase AI chips, marking the largest public commitment by a hyperscaler to diversify away from Nvidia dominance. The agreement includes a performance-based warrant for up to 160 million AMD shares—approximately 10% of the company—structured to vest alongside deployment milestones. The full stock award requires AMD’s share price to reach $600 (it closed Monday at $196.60).

The deal encompasses AMD’s MI540 series GPUs and latest-generation CPUs, with compute capacity sufficient to drive roughly six gigawatts of data center power demand. AMD CEO Lisa Su emphasized that “the CPU market is absolutely on fire” as AI inference workloads increasingly favor CPUs for their efficiency and ease of scaling. Meta CEO Mark Zuckerberg framed the partnership as “an important step” toward “personal superintelligence”—AI systems designed to deeply understand and empower individuals in their everyday lives.

The timing is strategic for both companies. For AMD, the Meta partnership provides the demand visibility and financial commitment needed to scale production and compete more directly with Nvidia in the AI accelerator market. For Meta, diversification reduces risk from Nvidia supply constraints and potentially improves pricing leverage. The company has pledged at least $600 billion in U.S. data center and AI infrastructure investments over the next several years, including $135 billion in capital expenditure for 2026 alone.

Industry analysts see the deal as validation that the AI infrastructure buildout remains on track despite recent market turbulence. However, Goldman Sachs Chief Economist Jan Hatzius offered a sobering counterpoint earlier in the week: AI investment contributed “basically zero” to U.S. GDP growth in 2025, largely because much of the equipment powering AI is imported. “A lot of the AI investment that we’re seeing in the U.S. adds to Taiwanese GDP, and it adds to Korean GDP but not really that much to U.S. GDP,” Hatzius said, highlighting the disconnect between AI investment narratives and economic reality.

OpenAI Enlists Big Four Consulting Firms to Push Enterprise AI Platform

Category: Enterprise and Adoption
Source: Fortune

OpenAI has announced partnerships with Boston Consulting Group, McKinsey & Co., Accenture, and Capgemini to help sell and implement its new Frontier AI agent platform, significantly expanding its enterprise go-to-market strategy. The consulting firms will help clients redesign workflows, integrate AI agents with existing software and systems, provide change management support, and offer industry-specific expertise OpenAI doesn’t have in-house.

Frontier, which OpenAI debuted earlier this month, is a system that allows businesses to build, deploy, supervise, and govern AI agents. The platform represents OpenAI’s effort to seize enterprise AI market momentum from rival Anthropic, which made substantial inroads over the past year with Claude Code and Claude Cowork. Under these new “Frontier Alliances,” each consulting firm is investing in dedicated practice groups and building teams certified on OpenAI technology, while OpenAI’s own “forward-deployed engineers” will work alongside consultants in client engagements.

BCG and McKinsey are positioned primarily as strategy and operating-model partners, helping leadership teams determine where and how to deploy agents at scale. Accenture and Capgemini take on more end-to-end systems integration roles, handling data architecture, cloud infrastructure, and the complex work of connecting Frontier to enterprises’ actual technology stacks. Bob Sternfels, McKinsey’s global managing partner, said CEOs must “rewire their businesses, reimagining domains and evolving how their people work” to capture value from agentic AI.

The announcement creates potential tension in the software industry ecosystem. The same consulting firms are deeply embedded with enterprise software vendors like Salesforce, Microsoft, and ServiceNow—companies whose products Frontier could potentially displace. For these SaaS vendors, having BCG and McKinsey actively evangelize an alternative platform to C-suite executives is an unwelcome development, particularly as investors have already punished their shares over concerns that customers will choose OpenAI’s and Anthropic’s agent platforms over traditional enterprise software.

Quick Updates

• Thermodynamic Computing Achieves Orders of Magnitude Energy Efficiency: Researchers generated images from noise using thermodynamic principles, requiring orders of magnitude less energy than current generative AI models. The approach represents a fundamental rethink of computing architecture that could dramatically reduce AI’s energy footprint. Source: Live Science

• Trump Administration Launches “Tech Corps” to Counter China: The U.S. has announced a new “tech corps” initiative aimed at spreading American AI influence globally and countering Chinese technology expansion. The program represents a Peace Corps-style approach to technology diplomacy. Source: CNBC

• India’s Sarvam Launches Multilingual AI Chat App: Indian AI startup Sarvam has launched Indus AI, a consumer chat application designed specifically for Indian languages, positioning itself as a domestic alternative in a market dominated by global technology companies. The app targets users beyond India’s English-speaking population with deeper regional language support. Source: TechPuts

• Anthropic Acquires Vercept to Advance Computer Use: Anthropic has acquired Vercept, a company specializing in AI perception and interaction, to accelerate development of Claude’s computer use capabilities. Vercept co-founders Kiana Ehsani, Luca Weihs, and Ross Girshick will join Anthropic as the company pushes toward agents that can operate within live applications. Source: Anthropic

• AI Coding Tools Reshape Software Economics: Tasks that once required significant time now execute quickly with tools like Claude Code, Cursor, and GitHub Copilot, reducing implementation friction but shifting competitive advantage to architecture, system design, and long-term maintainability rather than raw coding speed. Source: Reddit Discussion

• Ladybird Browser Adopts Rust with AI Assistance: The Ladybird browser project is rewriting portions of its JavaScript engine from C++ to Rust, with developer Andreas Kling using Claude Code and Codex to accelerate the translation. The port of 25,000 lines took about two weeks—work Kling estimates would have taken multiple months manually. Source: Ladybird.org

• Democrats Retreat from AI Data Center Courting: Potential 2028 presidential candidates including Illinois Gov. JB Pritzker, Pennsylvania Gov. Josh Shapiro, and Maryland Gov. Wes Moore are abruptly retreating from efforts to lure AI data centers, pivoting to protect voters from AI revolution consequences after previously offering lavish tax breaks. Source: Axios

• Anthropic Research Shows Software Development Agent Dominance: Nearly 50% of all agent tool calls through Anthropic’s public API are for software development, according to analysis of millions of real human-agent