Trump cancels AI safety executive order: Musk, Zuckerberg and Sacks have won
🔎 A phone call that derailed everything
In May 2026, Trump was supposed to sign a historic executive order imposing mandatory safety tests before the deployment of powerful AI models. Ceremony planned, tech executives invited, press releases prepared. Everything was locked in. Except that two billionaires — Elon Musk and Mark Zuckerberg — called the White House directly. Result: outright cancellation, with no rescheduled date.
This episode perfectly illustrates how regulatory power dilutes when the regulated have the regulator's personal number.
The key points
- Trump canceled the signing of an executive order imposing pre-deployment safety tests for powerful AI models, under direct pressure from Musk and Zuckerberg.
- David Sacks, the administration's "AI Czar", also pushed against the text he "hated", according to Axios.
- AI safety advocates in Washington are disappointed: it's a green light for the unfettered deployment of the most capable models like GPT-5.5, Grok 4.1 or Gemini 3.1 Pro.
- The paradox is total: Musk (xAI) and Zuckerberg (Superintelligence Lab) are both Trump's advisors and the primary beneficiaries of the lack of regulation.
Recommended tools
| Tool | Main use | Associated model | Ideal for |
|---|---|---|---|
| Hostinger | Web hosting + built-in AI | — | Deploying AI projects without friction |
What the torpedoed executive order planned
The order was to impose federal safety assessments before companies could deploy "frontier" AI models — those whose capabilities exceed a certain danger threshold. Specifically, independent tests on reliability, biases, and proliferation risks.
The idea was not new. The Biden administration had laid the initial groundwork with its October 2023 executive order, which used the Defense Production Act to force labs to share the results of their safety tests ("red-teaming") with the federal government.
Trump's text went further: mandatory tests before going to market, not just a posteriori reporting. For the models dominating the current rankings — GPT-5.5 (agentic score 98.2), Gemini 3 Pro Deep Think (95.4), Claude Opus 4.7 Adaptive (94.3) — this meant an additional validation process of several weeks, or even months.
It is precisely this delay that the industry refused.
Behind the scenes: Musk, Zuckerberg and the red phone
The story, revealed by Forbes and confirmed by The Independent, is remarkably straightforward. Musk and Zuckerberg called Trump to express their "concerns" — a euphemism for "we disagree and we want it scrapped."
According to The Guardian, the two CEOs warned the president that this order would harm the US economy and its competitive advantage in AI. The argument of competition with China was brandished as a rhetorical shield.
David Sacks, appointed "AI Czar" by Trump and a former partner of Peter Thiel, played a catalytic role. Axios reports that a source close to the matter claims Sacks "hated" the regulation. Sacks, who co-founded Craft Ventures and was an early investor in Facebook, is not a neutral player on the subject.
Trump himself justified the reversal with his usual phrasing: "I didn't like certain aspects," telling Politico that he was worried the EO "could have been a blocker." The signing, initially scheduled for a Thursday with several tech executives present, was simply canceled. No replacement date has been announced.
This sequence raises a fundamental question: when the main parties affected by a regulation are the regulator's official advisors, is the system designed to function?
The Musk-Zuckerberg paradox: judges and interested parties
Elon Musk has built part of his public communication around warning of the existential risks of AI. He co-founded OpenAI in 2015 partly for this reason, left the board when management took a more commercial turn, and launched xAI in 2023 with a decidedly "safety-first" discourse.
Except that in May 2026, xAI deploys Grok 4.1, ranked tied for 90th on the general benchmark and 14th in agentic. The model must compete with GPT-5.5 and Gemini 3.1 Pro. Each week of delay due to federal safety tests is a lost week against OpenAI and Google.
Musk knows this. And when he called Trump, he was defending xAI's interests, not those of AI safety. Ars Technica puts it bluntly: the tech industry lobbied against the order, fearing that testing would delay launches.
Zuckerberg is in an even more obvious position. Meta launched its Superintelligence Lab in early 2026, with a budget of several billion and a clear goal: catch up with OpenAI and Google. Meta's model does not appear in the current top 15, which means the lab is precisely in the catch-up phase where every month counts. Imposing safety tests at this stage means giving an additional advantage to the leaders.
These questions of security and ethics of personal AI avatars or the broader security of agentic systems like those described in Security and permissions in Hermes Agent become all the more critical when the regulatory framework is emptied of its substance.
Why safety advocates are sounding the alarm
Current agentic models are no longer chatbots. GPT-5.5 scores 98.2 on the agentic benchmark, meaning it can plan, execute multi-step tasks, correct its own errors, and interact with external tools almost autonomously.
Claude Opus 4.7 Adaptive (94.3) and Gemini 3 Pro Deep Think (95.4) are in the same zone. Moonshot AI's Kimi K2.6, in self-host, reaches 88.1 — a score that would have seemed like science fiction two years ago.
In this context, the absence of mandatory safety testing concretely means that:
No independent entity verifies a model's robustness before it is placed in the hands of millions of users. Companies do their own "red-teaming", which amounts to grading yourself on an exam.
Proliferation risks are real. A model like DeepSeek V4 Pro (88 overall, 84 in agentic high) is open-weight and can be deployed anywhere. Without a federal framework, nothing prevents a malicious actor from fine-tuning it for dangerous purposes.
Alignment scenarios remain independently unverified. When GPT-5.3 Codex (87 overall, 80 agentic) writes code almost autonomously, the question of whether it still follows safety guidelines in edge cases is decided solely by OpenAI.
In Washington, regulation advocates are described by Axios as "disappointed". That is an understatement. Trump's reversal is described by The Guardian as a "green light for unchecked tech power".
Biden vs Trump vs Europe : three radically different approaches
The contrast between regulatory approaches is striking.
The Biden approach (2023-2025): The October 2023 executive order used the Defense Production Act to require labs to share the results of their red-teaming with the government. No deployment block, but mandatory transparency. Insufficient according to safety advocates, but it was a first federal framework.
The Trump approach (2025-2026): Progressive dismantling. First, the revocation of the Biden EO in January 2025, then the aborted attempt to replace it with a new text in May 2026 — torpedoed from the inside. The message is clear: the market regulates itself.
The European approach (AI Act): The regulation now in force classifies AI systems by risk level. "Frontier" (general-purpose) models are subject to transparency obligations, technical documentation, and systemic risk evaluation requirements before being placed on the market. Sanctions can reach up to 7% of global revenue.
| Approach | Mechanism | Possible block? | Sanctions |
|---|---|---|---|
| Biden EO (2023) | Mandatory transparency via DPA | No | No formal ones |
| Trump EO (May 2026) | Cancelled before signature | — | — |
| AI Act (EU) | Classification by risk + pre-deployment evaluation | Yes, for unacceptable risk models | Up to 7% global revenue |
The risk for American companies is twofold. On the one hand, they no longer have federal constraints. On the other hand, if they want to operate in Europe — which is the case for all of them — they still have to comply with the AI Act. Except that the tests required by Brussels are not the same as those Washington might have imposed, and the lack of harmonization creates a costly regulatory fragmentation.
China's argument: valid or a smokescreen?
The main argument from Musk and Zuckerberg, as reported by The Independent, is that mandatory safety testing would slow down the US in the AI race against China.
This argument holds some truth. China is deploying competitive models: DeepSeek V4 Pro (88 overall) and Z.AI's GLM-5.1 (83) are in the top 15. Moonshot AI's Kimi K2.6 reaches 84 overall and 88.1 in agentic self-host. Beijing does not impose independent safety tests equivalent to those being discussed in Washington.
But the argument also has major flaws.
First, safety testing does not mean months of delays. Red-teams already exist in all labs. The idea was to make them verifiable by a third party, not to create them from scratch.
Second, safety is a competitive advantage. A model that does not make catastrophic errors is a model that businesses adopt more easily. Tools that allow you to earn 300€/month without coding thanks to AI are viable precisely because the underlying models are reliable enough.
Finally, the China argument is a false promise when wielded by CEOs whose models are not yet at the level of the leaders. Slowing down GPT-5.5 and Gemini 3.1 Pro with testing would have helped challengers like xAI and Meta, not the other way around.
What model scores tell us about risks
The model landscape in June 2025 shows an acceleration that makes the safety question even more pressing.
In general-purpose models, the gap between the first (Gemini 3.1 Pro, 92) and the 15th is only 9 points. The compression of scores means that "average" models are becoming highly capable very quickly. GLM-5.1 (83) and Claude Sonnet 4.6 (83) offer performance that is sufficient for the majority of use cases.
In agentic models, the concentration is even more striking. GPT-5.5 dominates at 98.2, but GPT-5.4 Pro (91.8) and Claude Opus 4.7 (94.3) are in a zone where operational autonomy is real. An agent that plans, executes, iterates, and corrects itself with a score of 94+ is no longer an assistant — it is an autonomous collaborator.
When these models are connected to real tools (files, APIs, databases, servers), the consequences of an alignment failure are no longer theoretical. The question of permissions and security in agentic systems, such as the one addressed in Sécurité et permissions dans Hermes Agent, becomes an immediate operational issue.
Without a federal testing framework, the only line of defense is the internal red-teaming of each lab. OpenAI, Anthropic, and Google have strong safety teams. But their work is not audited by a third party, and commercial incentives push for speed, not caution.
The role of David Sacks: an anti-regulation AI Czar
David Sacks deserves special attention. Appointed "AI Czar" by Trump in 2025, he is theoretically the head of the administration's AI policy. His position on this matter is revealing.
Axios reports that Sacks "hated" the executive order. It is striking: the person in charge of steering the country's AI policy actively opposes the only serious safety measure discussed under this presidency.
Sacks is not a neutral technologist. He is a VC investor (Craft Ventures), former COO of PayPal, and a long-time friend of Musk and Peter Thiel. His network is exactly the people who benefit from the absence of regulation. Appointing him AI Czar and asking him to regulate his friends and partners is like appointing an oil industry lobbyist to head the EPA.
According to The Guardian, Sacks was "involved" in lobbying against the EO. His role was not passive: he actively worked to derail the text from within the administration.
The message sent to safety advocates is brutal: the White House door is closed.
"Anti-doomer" rhetoric as a political weapon
An interesting element of this affair is the vocabulary used. Axios reports that "anti-doomer" feedback derailed the EO. The term "doomer" has become a rhetorical weapon to discredit anyone who expresses concerns about AI risks.
This is no accident. "Anti-doomer" discourse serves specific interests: it allows for framing AI safety as an obsession of paranoids, not as a requirement of serious engineering. Yet, safety researchers are not "doomers". They are engineers studying real problems: how to ensure that an agentic model with a score of 98 doesn't make destructive decisions when deployed at scale.
The paradox is that Musk himself was one of the biggest promoters of the "AI is an existential risk" discourse for years. His turnaround in May 2026 is not a change in scientific analysis, it is a change in economic position: xAI now has models to sell.
And now? Post-cancellation scenarios
The cancellation of this EO leaves a regulatory void that will likely not be filled for a long time. Three scenarios are possible.
Scenario 1: the unfettered status quo. No new federal text. Labs deploy freely, do their own red-teaming, and the market penalizes failing models. This is the preferred scenario of Musk, Zuckerberg, and Sacks. The risk: a major incident (mass deepfake, AI agent making destructive decisions) that triggers a brutal and poorly calibrated regulatory reaction.
Scenario 2: Congress steps in. Senators from both parties (they exist, even among Republicans) could introduce a bill. But Congress is slow, divided, and tech lobbyists are well entrenched. Low probability in the medium term.
Scenario 3: fragmentation. Europe applies its AI Act, California strengthens its own laws (like SB 1047 which nearly passed in 2024), and the United States has no coherent federal framework. American companies end up with a chaotic regulatory patchwork. This is probably the most realistic scenario.
In any case, the May 2026 cancellation marks a tipping point. The window for establishing an ambitious federal AI safety framework is closing.
❌ Common mistakes
Mistake 1: Confusing AI safety and "anti-innovation"
AI safety is not a brake on innovation, it's reliability engineering. Safety testing for aviation didn't kill the aerospace industry — it made it trustworthy. The parallel is direct.
Mistake 2: Thinking that labs regulate themselves
History shows that self-regulation works when there is a threat of credible external regulation. Without this threat, the incentives to cut corners are too strong. Internal red-teaming is essential but not sufficient.
Mistake 3: Believing that Musk is a safety advocate
Musk was one, when it suited his position. In 2026, xAI has models to deploy and investors to satisfy. His actions (calling Trump to torpedo the EO) speak louder than his past statements.
❓ Frequently Asked Questions
Which models were affected by this executive order?
The EO targeted "frontier" models, meaning the most capable ones. In practice, GPT-5.5 (98.2 agentic), Gemini 3 Pro Deep Think (95.4), Claude Opus 4.7 (94.3) and their successors would have been the first affected.
Does the European AI Act apply to American companies?
Yes. Any model deployed in the EU, regardless of the company's origin, is subject to the AI Act. American labs therefore still have to conduct risk assessments — but according to the Brussels framework, not Washington's.
Why did Trump cave so easily?
Trump "hates regulation" according to a source cited by Axios. When two of his closest tech allies (Musk and Zuckerberg) and his own AI Czar (Sacks) tell him not to sign, he doesn't need much persuasion.
Will there be a new replacement text?
No date has been announced. Politico indicates that it is "not clear" when a signing could be rescheduled. In Washington parlance, that probably means never in this form.
✅ Conclusion
The cancellation of the May 2026 executive order is not a mere bump in the road, it is the result of effective lobbying carried out by players who are literally sitting at the table of power. Musk, Zuckerberg and Sacks won this battle. The question is who will lose the next one.