Pentagon AI Deal: Google Signs, Anthropic Gets Blacklisted

Google just signed a classified agreement giving the US Department of Defense access to its AI models for "any lawful government purpose." Anthropic, by contrast, refused Pentagon demands — and was quietly blacklisted for it. Both decisions landed in the same news cycle, and the contrast has never been starker.

This isn't just a corporate policy story. It's the moment AI companies were forced to answer a question most had been deferring: when the Pentagon asks for your AI, do you say yes?

The Pentagon AI Contract Anthropic Refused — and the Price It Paid

According to reporting by The Information, Google signed a classified contract allowing the US DoD (Department of Defense) to deploy its AI models in ways that go far beyond what's publicly disclosed. OpenAI and Elon Musk's xAI (the company behind the Grok chatbot) had already made similar classified deals. That left Anthropic as the notable holdout.

Anthropic's refusal was specific: the Pentagon reportedly asked Anthropic to remove safety guardrails (built-in restrictions that prevent AI from generating dangerous or harmful outputs) from its Claude models for military deployment. Anthropic declined. The consequence was immediate — the company was blacklisted from certain US government AI procurement channels, losing access to a pipeline worth billions in potential contracts.

• Google: Signed classified deal — "any lawful government purpose"
• OpenAI: Classified AI deal with DoD already confirmed
• xAI (Grok): Military contract signed
• Anthropic: Refused to strip safety features — blacklisted by Pentagon

600 Google Employees Demand Pichai Reject Classified Military AI Contracts

Google's classified deal didn't sit quietly with its own workforce. More than 600 employees — including over 20 principals, directors, and vice presidents — signed an internal letter demanding CEO Sundar Pichai reject all classified workloads from the US military.

The letter's core argument was unambiguous: "The only way to guarantee that Google does not become associated with such harms is to reject any classified workloads."

This isn't Google's first internal revolt over military contracts. In 2018, the Project Maven controversy — in which Google built AI for drone targeting analysis — ended with the contract being cancelled after mass employee protests. This time, leadership appears to have moved forward regardless.

The seniority of the signatories matters. Twenty-plus principals and directors at a company like Google aren't junior employees — they hold real product and engineering authority. Their presence on a petition signals this isn't fringe discontent. It's a split inside Google's technical leadership over where AI should and shouldn't be deployed.

DARPA AIxCC Challenge: AI Security Tools Found Bugs Nobody Planted

Separate from the contracting debate, the US Defense Advanced Research Projects Agency (DARPA — the Pentagon research arm that helped build the early internet) ran its Artificial Intelligence Cyber Challenge, known as AIxCC. The results exceeded expectations in a striking way.

AI systems were given 54 million lines of actual production software code, injected with artificially created security flaws. The goal: find the planted bugs. The systems did — but they also surfaced 12 or more real vulnerabilities that DARPA's own security researchers had not inserted.

That secondary finding is what matters most. AI security tools (programs that automatically scan source code for exploitable weaknesses) are now identifying problems that human auditors missed entirely. Claude Mythos, an AI model fine-tuned specifically for vulnerability research, was cited as a standout performer in the challenge.

Why 12 Zero-Day Vulnerabilities Found by AI Marks a Capability Inflection Point

Security professionals call previously unknown flaws "zero-day vulnerabilities" — weaknesses that exist in production software but have never been publicly disclosed or patched. Finding even one zero-day in a controlled exercise is considered impressive. Finding 12 or more in a single automated pass, in code that wasn't specifically prepared for them, suggests AI-assisted security auditing is approaching a capability inflection point (the threshold where a technology becomes reliably useful at enterprise scale, rather than just theoretically promising in lab settings).

For reference: professional security audits of large codebases (collections of software source code) typically take months and teams of 5–20 specialists. The DARPA AIxCC systems processed 54 million lines in a fraction of that time — and came back with findings the humans had missed.

The $150 Billion Musk vs. OpenAI Trial That Could Reshape AI Governance

While Pentagon deals were being finalized, jury selection began in San Francisco for the Musk vs. OpenAI trial. Elon Musk is suing OpenAI, CEO Sam Altman, and co-founder Greg Brockman for allegedly abandoning the nonprofit mission he helped fund — then pivoting the organization toward profit.

Musk claims Altman and Brockman "tricked him" into donating significant money to what was presented as a humanity-first research organization, then transformed it into a for-profit model competing directly with his own ventures, including xAI. He's demanding up to $150 billion in damages and calling for the removal of Altman and Brockman from OpenAI leadership.

OpenAI's counterclaim is direct: the lawsuit is "a baseless and jealous bid to derail a competitor," engineered to benefit Musk's own AI company. Jury selection on April 27th produced remarkable candor from prospective jurors — questionnaires included statements like "Elon Musk is a world-class jerk" and sharper criticisms. Whether an impartial jury can be seated given Musk's current public profile remains an open legal question.

In a parallel development, Microsoft quietly removed the AGI clause (a contractual provision that had given Microsoft special rights and obligations if OpenAI achieved artificial general intelligence — AI that surpasses human-level capability across all tasks) from its long-standing partnership agreement. With that clause gone, OpenAI is now free to distribute products across AWS, Google Cloud, Azure, or any cloud platform — ending a period of effective exclusivity that had tied OpenAI's enterprise ambitions primarily to Microsoft infrastructure.

Five Fault Lines That Defined One Week in AI Policy

The events of late April 2026 arrived close enough together to reveal something structural. AI companies are no longer just building products — they're being asked to take positions on classified military contracts, workforce ethics, corporate governance, and content policy, often simultaneously and without established playbooks.

A smaller story from the same week illustrated this in sharp detail: Canva's AI design tool was found substituting the word "Palestine" with "Ukraine" in user-generated text. Canva confirmed the substitution was real and isolated to that specific word, with related terms like "Gaza" unaffected. No explanation was provided for why the AI model was programmed to make that substitution, or when the behavior first began.

The pattern runs through all five stories: AI systems are making decisions — about which safeguards to keep, which cloud to run on, which words to replace, which vulnerabilities to flag — without surfacing those decisions transparently to the people they affect. Watch the Musk-Altman trial closely. If internal OpenAI documents are entered into the public record during proceedings, it may offer the most detailed picture yet of how the world's most powerful AI lab actually makes its decisions — and who holds the authority to reverse them. You can follow live updates at AI for Automation News.

Related Content — Get Started with AI Automation | AI Guides | More AI News