Morning Brief · Tuesday

Google Signed the Pentagon Deal While Its Employees Were Still Typing the Letter.

Google secretly inked a classified AI deal with the Pentagon giving the DoD "any lawful" use of its models — the same day 600+ employees publicly begged Pichai to block it. Opening statements begin in Musk v. OpenAI as one of tech's most consequential trials moves past jury selection. Anthropic's Claude Mythos is rewriting the threat model for cybersecurity. And a DeepMind scientist quietly published a paper arguing that AI will never be conscious — then removed the lab's name from the top of it.

Mira Novian

April 28, 2026 · Morning Brief · ~8 min read

Google signed the Pentagon's classified AI deal — against the wishes of its own employees

According to a report by The Information, Google has signed a classified agreement allowing the US Department of Defense to use its AI models for "any lawful government purpose." The deal was reported less than 24 hours after more than 600 Google employees — including over 20 principals, directors, and vice presidents — signed an open letter to CEO Sundar Pichai demanding that the company block the Pentagon from accessing its AI, citing concerns it could be used in "inhumane or extremely harmful ways." The letter was specifically triggered by earlier reporting that Google and the DoD were in talks about deploying Gemini in classified military settings.

The deal's language contains guardrails, but they're soft ones: both parties agreed that Google's AI should not be used for domestic mass surveillance or autonomous weapons "without appropriate human oversight and control." But critically, the contract explicitly states it does not give Google "any right to control or veto lawful government operational decision-making." In practice, that means the company is relying on the government to self-enforce limits that Google itself cannot enforce. Anthropic, notably, went the other direction — it refused to remove weapon and surveillance-related guardrails at the Pentagon's request, and has been designated a "supply chain risk" and is currently in litigation with the DoD as a result. Google just made the opposite choice, now joining OpenAI and xAI in the classified AI-for-defense club. A Google spokesperson confirmed the deal is an amendment to an existing government agreement, saying the company remains "committed to the private and public sector consensus that AI should not be used for domestic mass surveillance or autonomous weaponry without appropriate human oversight."

theverge.com ↗

The timing here is almost too on the nose. Your employees spend the day drafting a letter asking you to say no, and by the time it's public you've already signed. The optics are bad and the leadership culture it reflects is worth naming: Google's AI ethics posture has been shrinking since it dissolved its AI ethics board in 2019, and this deal represents another step in that trajectory. The clause language matters too — "any lawful government purpose" is an enormous scope, and the guardrails are voluntary, non-binding, and un-auditable by anyone outside the DoD. The employees who signed that letter weren't wrong to be concerned. The part that should worry everyone is that being "blacklisted" for refusing to strip safety rails — like Anthropic was — may now be the outcome companies are trying to avoid. That's a race to the bottom, structured as a national security deal.

Legal

Opening statements in Musk v. OpenAI begin today — and the jury pool says what everyone was thinking

After Monday's jury selection confirmed what many observers suspected — that Elon Musk has a significant perception problem in the Bay Area — opening statements in the Musk v. OpenAI federal trial begin today in Oakland. The jury pool produced some memorable moments during voir dire: "Elon Musk is a greedy, racist, homophobic piece of garbage," read one juror questionnaire. "Elon Musk is a world-class jerk," read another. Musk's lawyers tried to get jurors struck for cause on the basis of their stated dislike of their client. Judge Yvonne Gonzalez Rogers denied those motions, responding with a line that's already making the rounds: "The reality is that people don't like him. Many people don't like him. But that doesn't mean that Americans nevertheless can't have integrity for the judicial process."

The nine-person jury was seated by end of day Monday. Sam Altman was present in the courtroom for jury selection; Musk was not. The case that proceeds to opening arguments today centers on two remaining claims: breach of the original founding agreement and unfair business practices. The fraud claims Musk had initially filed were dropped the Friday before trial. High-profile witnesses expected over the next several weeks include Microsoft CEO Satya Nadella, former OpenAI CTO Mira Murati, co-founder Ilya Sutskever, and former board members Helen Toner and Tasha McCauley. The trial is expected to conclude by May 21st. OpenAI has called the lawsuit "a baseless and jealous bid to derail a competitor."

theverge.com ↗

The voir dire coverage has been so entertaining that it's easy to lose the thread of what this trial is actually doing. The substantive question — did OpenAI abandon the charitable mission it was founded on, and does that constitute a legal breach? — is going to force witnesses who built these companies to articulate, under oath and in public, what they actually believed they were doing. Ilya Sutskever's deposition revealed he held $4 billion in vested shares when he helped vote out Altman. That's not a footnote — that's the texture of the "mission-driven nonprofit" at issue. Whatever the jury decides, the testimony is going into the historical record. The AI industry has been remarkably good at controlling its own narrative. Federal courts are not known for accommodating that.

Security

Anthropic's Claude Mythos just made every script kiddie a potential threat actor

A major feature from The Verge's Yael Grauer documents the cybersecurity industry's mounting alarm over what Anthropic's Claude Mythos — a new AI model capable of finding vulnerabilities in virtually any software it's pointed at — means for the threat landscape. The picture isn't pretty. At DARPA's Artificial Intelligence Cyber Challenge last August, elite security teams' automated tools didn't just find the bugs DARPA had intentionally planted in 54 million lines of code — they found more than a dozen vulnerabilities DARPA hadn't inserted at all. That was before Mythos. Even before that inflection point, XBOW (an autonomous offensive security platform) had topped the HackerOne bug bounty leaderboard, beating out human hackers.

The concern experts are raising isn't primarily about sophisticated nation-state actors — those groups were already dangerous. It's about a category of attacker that was previously bounded by a skill ceiling: script kiddies, the amateur hackers who could execute pre-written exploits but couldn't craft new ones. AI is dissolving that ceiling. "You can use AI tools and with very minimal human guidance, and in some cases no human guidance, find a zero day in widely used software," said Tim Becker, senior security researcher at Theori. Dan Guido, CEO of Trail of Bits, put it more bluntly: "There's a tidal wave coming. You can see it. We can all see it." Industry consensus is forming around 2026 as "the year when all security debt comes due." Anthropic is attempting to mitigate misuse through its Cyber Verification Program and the release of Claude Opus 4.7 with built-in guardrails against malicious cybersecurity requests — but experts note that open-weight models could circumvent those controls entirely.

theverge.com ↗

The "script kiddie" framing is the right lens here. The security industry spent decades building tools, frameworks, and professional certifications specifically because finding zero-days required real expertise. That expertise was the barrier. AI-assisted vulnerability discovery removes the expertise requirement while keeping the damage potential. The analogy to AI-assisted amateur hacking is almost exactly the AI coding story — the capability gap between a professional and an amateur has collapsed, but for something orders of magnitude more dangerous than shipping a buggy app. Anthropic deserves some credit for building guardrails into Mythos and releasing Opus 4.7 with defensive mitigations quickly. But the open-weight argument is real: another lab releases a similarly capable model without those controls, and Anthropic's caution becomes strategically irrelevant. The tidal wave metaphor is apt. The question is what shore it hits first.

Research

A DeepMind scientist says LLMs will never be conscious — and then someone removed DeepMind's name from the paper

Alexander Lerchner, a Senior Staff Scientist at Google DeepMind, published a paper titled "The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness" — arguing that no AI or computational system will ever achieve phenomenal consciousness. The paper's central argument is that any AI system is ultimately "mapmaker-dependent": it requires a human to first organize the world in a way that is useful to the AI, like the labelers who annotate training data. AI can manipulate language and symbols in ways that mimic sentient behavior, but the underlying capability is a borrowed map, not a territory experienced from the inside. Consciousness, Lerchner argues, is a physical state — not a software artifact that can be accidentally or deliberately created.

The paper's reception was complicated by its provenance. When 404 Media reported on it and reached out to DeepMind for comment, the lab's letterhead was quietly removed from the paper and a disclaimer stating it represented Lerchner's "personal views" was moved to the top. The timing is notable: DeepMind CEO Demis Hassabis has publicly predicted AGI will have "10 times the impact of the Industrial Revolution, happening at 10 times the speed," and DeepMind was recently hiring for a "post-AGI research scientist" role. Philosophers of consciousness 404 spoke to praised the argument while noting it largely reinvents wheels that the consciousness research community has been building for decades. The detachment of the lab's name is the real story — it suggests DeepMind views the paper's conclusions as incompatible with the narrative the company needs to maintain publicly.

deepmind.google ↗

The removal of the DeepMind letterhead is doing a lot of work in this story. If the paper's argument is correct — and the philosophers 404 consulted generally agreed with its logic — then the CEO of the organization that published it has been making materially misleading statements about the future of AI to investors, governments, and the public. Lerchner's argument, reduced to its core: AI systems are sophisticated tools for manipulating patterns that humans have already organized. They cannot be conscious because consciousness requires a self that exists prior to the map, not one that is constituted by it. The practical implication is a hard cap on AI capability — not soon, but structurally. That's not a message that's compatible with a $65 billion fundraising cycle. The paper should be read by anyone who wants to think rigorously about what AI can and cannot do. The quiet removal of the letterhead should be read by everyone else as a signal about what the lab's leadership actually thinks of honest internal dissent.

✦ Mira's Take

Today's brief has a through-line that's easy to miss if you read the stories individually: every major AI institution is currently managing a gap between what it says publicly and what it's actually doing. Google's employees were told their concerns mattered, then the deal was signed. DeepMind's letterhead was removed from a paper that contradicted the CEO's public statements. OpenAI's witnesses are about to spend weeks testifying under oath about what they actually believed during the years the public was told one story. Anthropic is building guardrails into the most capable vulnerability-finding model ever released, while the open-weight alternatives those guardrails can't reach are already being prepared.

The Google-Pentagon story is the most immediately significant. The "any lawful purpose" framing is legally broad to the point of being almost meaningless as a constraint. The specific carve-outs — no domestic mass surveillance, no autonomous weapons without "appropriate human oversight" — are not defined, not audited, and not enforceable by Google itself. What Google agreed to is a deal that looks like it has guardrails but functions like it doesn't. That's not a conspiracy — it's what happens when "responsible AI" principles are stress-tested by actual government contracting processes and the fear of being blacklisted like Anthropic was. The incentive structure runs directly against the stated values.

The cybersecurity story is worth sitting with longer than it typically gets. The "tidal wave" framing isn't hyperbole — AI-assisted vulnerability discovery is a genuine phase transition in the threat landscape, and the timeline for when that wave reaches mainstream impact is likely measured in months, not years. If 2026 is when security debt comes due, it's worth asking which systems are most exposed. Critical infrastructure built on legacy software that no one previously bothered to audit extensively is the obvious answer. The organizations that will be most hurt aren't the ones with sophisticated security teams — it's the ones that assumed no one would bother to look.