Elon Musk used the stage at the All-In Summit in Los Angeles on September 15, 2026, to pitch a fix for the AI industry’s trust problem: stop letting labs grade their own homework, and make rivals test each other’s models before anyone ships. The remarks, first reported by CNBC and picked up within hours by The News International, landed at a moment when frontier AI releases are colliding with mounting public anxiety about how fast the technology is moving.
Musk’s proposal is simple to state and hard to execute: xAI, OpenAI, Anthropic, Google, Meta, and what he called “three or four” leading Chinese AI companies would each run their internal safety test suites, known as test harnesses, against every other lab’s unreleased frontier model. Instead of a single company deciding on its own whether a system is safe to launch, the model would sit in front of a panel of rivals who have every incentive to find something wrong with it.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Musk actually proposed at the All-In Summit
According to the CNBC account of the panel, Musk said, “I think it would be wise to do as soon as possible, if not immediately, would be to have the major AI competitors test each other’s models.” He framed the plan around the safety infrastructure labs already have in place. Per AI Weekly’s roundup of the appearance, Musk described the mechanism directly: “All the AI companies have a test harness, a series of tests that you give to any model to see if it’s going to build bioweapons or nuclear bombs or be deliberately deceptive.” His pitch was to cross-wire those harnesses, so that Anthropic’s safety tests run against xAI’s models, OpenAI’s tests run against Google’s models, and so on down the roster of frontier labs.
The reasoning Musk gave for why this beats the current setup was blunt. As CNBC quoted him, “Instead of grading your own homework, you would at least have competitors grading your homework and raising the alarm if they see concerns.” That framing matters because it reframes AI safety testing from an internal compliance exercise into something closer to adversarial peer review, the kind used in academic publishing or financial auditing, where the reviewer has no stake in the outcome looking good.
This is not a brand-new idea from Musk. He floated an earlier, less detailed version of it in a July 2026 interview with The Economist, reported by TeslaNorth’s coverage of the September remarks, where he suggested leading labs hold informal calls every one to two weeks and give competitors a one-to-two-week early-access window before shipping a major new model. The September comments at All-In expanded that framework, adding the explicit cross-testing mechanic and, notably, floating Chinese AI companies as participants alongside the usual US players.
Why this story matters right now
The timing is not incidental. Musk’s remarks arrived in the same week that OpenAI CEO Sam Altman was apologizing for a rocky GPT-6 Astra rollout that locked out paying subscribers, a story The News International covered directly alongside its report on Musk’s peer-review idea. Frontier AI releases have been shipping faster and drawing more scrutiny simultaneously, and Musk’s own AI lab is racing to keep up: he has separately teased that xAI’s unreleased Grok 4.7 sits roughly on par with Anthropic’s Claude Opus 5.0 but trails the newer Claude Opus 5.1, according to Cryptopolitan’s reporting on his AGI comments tied to a planned Grok 5. A closer look at how the Grok lineup compares to rival chips and models shows just how tight that race has become.
That context is important for reading the proposal correctly. Musk is not a neutral observer calling for oversight of an industry he is not part of. He runs xAI, one of the six or seven labs that would be subject to the exact scrutiny he’s proposing. His plan effectively asks Anthropic, OpenAI, and Google to open a testing window into Grok before it ships, in exchange for the same access into their own unreleased systems. Whether that trade is appealing depends entirely on how much each lab believes it has to gain versus lose from letting a commercial rival poke at a model before launch.
The mechanics: test harnesses, API access, and cadence
Reporting on the proposal, including coverage from IBTimes and ChainCatcher, lays out three moving parts to how this would actually function if labs agreed to it.
First, the test harness itself. Every major lab maintains an internal battery of evaluations that probe a model for dangerous capabilities before release, including whether it can meaningfully assist with building biological or nuclear weapons, and whether it engages in deceptive or manipulative behavior toward users. These harnesses are proprietary today. Musk’s proposal would have each lab run its own harness against a competitor’s model rather than only its own, creating overlapping coverage that a single company’s internal team might miss.
Second, pre-release API access. For cross-testing to work, a lab would need to grant a rival API access to an unreleased model, not the underlying source code or weights, but enough functional access to run adversarial prompts and safety evaluations against it. Musk described this directly in comments reported by AI Weekly: “Basically, it means providing API access before the model is officially released.”
Third, a recurring cadence. Musk’s July 2026 comments to The Economist proposed informal calls between labs every one to two weeks, paired with a one-to-two-week early-access window for any major new model before public launch. That is a meaningfully short runway for a competitor to identify a dangerous capability, and it puts a real ceiling on how deep any cross-lab review could realistically go before a scheduled release date arrives.
How rivals have responded so far
Formal, on-record commitments from other labs have not materialized as of this writing. What has surfaced is more indirect. Musk referenced Anthropic CEO Dario Amodei’s public calls for stronger AI oversight, writing in a post that “Dario is right that there should be some oversight. Peer review of AI by competitors is the right way to start this off.” That is the clearest signal of alignment between two lab leaders on record right now, though it stops short of Anthropic committing to the specific cross-testing mechanism Musk described.
Musk also drew a contrast between labs during the All-In appearance, telling the audience that Anthropic puts more care into AI safety than OpenAI does, according to the trending-topic summary compiled around the panel. That comment reads as much like industry positioning as it does a safety endorsement. OpenAI, Google, and Meta have not issued public statements responding to the cross-testing proposal specifically, and no Chinese AI company has commented on Musk’s suggestion that “three or four” of them join the framework.
The China dimension
The most striking part of the September version of Musk’s pitch is that he explicitly wants Chinese AI companies inside the testing framework, not just US labs. The News International quoted him proposing exactly that: “We should probably get an agreement with China. It would be peer review, where the leading AI companies all test each other’s models before release.”
That is a heavier lift than getting OpenAI and Anthropic to share API access with each other. US and Chinese AI labs operate under different export-control regimes, different data-sovereignty rules, and a broader geopolitical rivalry over AI leadership that has already produced friction this year, including Chinese officials publicly rejecting US claims about AI model distillation. Asking labs on opposite sides of that rivalry to hand each other pre-release access to frontier models, even just through an API, runs into national-security review processes that go well beyond what any single CEO can negotiate through a summit panel or a social media post.
Competitive comparison: the labs Musk wants in the room
Musk named a specific set of participants during his remarks: xAI, OpenAI, Anthropic, Google, Meta, and several unnamed Chinese firms. Each of those companies has a different amount to gain or lose from the arrangement, based on where their models currently sit competitively. For a fuller breakdown of how these systems stack up on benchmarks and pricing, see this GPT-6 Astra vs Claude Opus 5 vs Gemini 3.8 Flash comparison, and the broader guide to the best AI models of 2026.
| Lab | Flagship model referenced | Stated safety posture | Public stance on cross-testing |
|---|---|---|---|
| xAI | Grok 4.7 (unreleased), Grok 5 teased | Musk is the proposal’s author | Actively proposing the framework |
| OpenAI | GPT-6 Astra | Faced criticism over Astra rollout issues | No public response reported |
| Anthropic | Claude Opus 5.0 / Opus 5.1 | Praised by Musk as more safety-focused | Amodei’s oversight remarks referenced approvingly by Musk |
| Gemini 3.8 family | Not directly addressed in Musk’s remarks | No public response reported | |
| Meta | Not specified by Musk | Not directly addressed in Musk’s remarks | No public response reported |
| Chinese AI firms (unnamed, “three or four”) | Not specified by Musk | Not directly addressed in Musk’s remarks | No public response reported |
Historical context: why self-policing is under pressure
Musk’s cross-testing idea did not emerge in a vacuum. It follows a string of 2026 stories about AI labs facing pressure to slow down or add guardrails to how fast frontier models ship. OpenAI and Anthropic executives have publicly discussed the idea of a development “brake” earlier this year, and Musk himself has pushed back on parts of that broader slowdown conversation in other contexts, including rejecting a specific slowdown deal floated elsewhere in the industry over what he characterized as an acceptable level of risk.
The consistent thread across these episodes is a disagreement not over whether AI development carries risk, but over the right mechanism for managing it. Some in the industry favor government regulation. Others favor voluntary pauses. Musk’s peer-review pitch is notable because it tries to route around both of those options, betting instead that competitive and reputational pressure, what he described as the risk of “egg on face” and potential legal liability if a lab ignores a rival’s warning and harm follows, will do more to keep frontier releases safe than a regulator or a self-imposed moratorium would.
Where the plan could break down
Even setting aside whether rival labs are willing to participate, the mechanics of Musk’s proposal carry real friction points. Handing a competitor pre-release API access to an unreleased frontier model, even without exposing weights or training data, creates exposure to intellectual-property leakage. A rival running thousands of adversarial prompts against an unreleased model can learn a great deal about its capabilities, guardrails, and failure modes, information with direct commercial value in a market where frontier labs are racing each other on benchmarks and pricing.
There’s also the question of incentive alignment during the review itself. A lab asked to test a rival’s unreleased model has a commercial interest in that model launching late, launching with a damaged reputation, or not launching at all. Musk’s framework assumes labs will act as neutral safety reviewers even when they are also direct competitors for the same enterprise customers and API revenue. Reports on the proposal note that no rival lab has formally agreed to adopt the mechanism yet, which suggests other executives may be weighing exactly this tension.
Finally, the proposed one-to-two-week access window, drawn from Musk’s July 2026 comments, is short relative to the scope of what the harnesses are meant to catch. Bioweapon-assistance testing, deception probing, and broader red-teaming for frontier models typically take safety teams considerably longer when done thoroughly inside a single lab. Compressing that into a window shared with an external, competing organization raises questions about how deep the review could realistically go before a launch date arrives regardless of what testing turns up.
Market and industry impact
For now, the proposal is a talking point rather than an operating agreement, so it has not moved product roadmaps or funding decisions. But the framing carries weight because of who is saying it. Musk runs one of the handful of companies with the compute and talent to compete at the frontier, and his public comments on AI safety tend to shape how policymakers and the press cover the sector, whether or not other labs act on them.
If even two or three labs adopted a limited version of cross-testing, that would set a de facto industry norm that regulators in the US, EU, and China could later point to when drafting formal AI safety requirements. Conversely, if the idea stalls the way earlier calls for a development pause largely have, it becomes one more data point supporting the argument that voluntary self-regulation among frontier labs consistently loses out to competitive pressure to ship.
Predictions: what happens next
- Expect Anthropic to be the lab most likely to publicly engage with some version of Musk’s proposal, given Dario Amodei’s existing public advocacy for AI oversight and Musk’s own comments crediting Anthropic’s safety focus.
- OpenAI and Google are unlikely to commit to full API-level cross-testing in the near term, given the commercial sensitivity of unreleased frontier models and the lack of any public response so far.
- A Chinese-lab agreement, if it happens at all, will likely take far longer than a US-only arrangement given export-control and national-security review requirements on both sides.
- Watch for narrower, bilateral pilot versions of the idea, such as two labs agreeing to a limited safety-test exchange, rather than the full six-or-more-lab framework Musk described.
- Regulators in the US and EU will likely cite the proposal in upcoming AI-safety policy discussions as evidence that industry itself sees a gap in current self-testing practices, regardless of whether any lab actually adopts it.
What a shared test harness would need to look like
For engineers and researchers tracking this story, the technical shape of a cross-lab test harness is worth spelling out, since it’s the part of Musk’s proposal most likely to actually get built in some form even if the full multi-lab agreement never materializes.
| Component | Purpose | Current status per reporting |
|---|---|---|
| Bioweapon-assistance evaluation suite | Tests whether a model can meaningfully assist with biological weapons development | Already exists internally at major labs, per Musk’s description |
| Nuclear-assistance evaluation suite | Tests whether a model can meaningfully assist with nuclear weapons development | Already exists internally at major labs, per Musk’s description |
| Deception and manipulation probing | Tests whether a model behaves deceptively toward users or evaluators | Already exists internally at major labs, per Musk’s description |
| Cross-lab API access window | Grants a rival lab functional access to run its harness against an unreleased model | Proposed only; not yet adopted by any lab |
| Recurring safety call cadence | Weekly or biweekly coordination between labs on emerging concerns | Proposed only; not yet adopted by any lab |
Notably, none of the five components above require new technology. Every major lab already has evaluation suites covering the risk categories Musk named. What’s missing is the access agreement, the legal framework around IP exposure and liability, and the willingness of commercial rivals to open a pre-release window to each other. That is a governance and business-terms problem more than an engineering one, which is part of why it’s proven easier to propose at a summit than to formalize into a binding industry agreement.
How this compares to existing AI safety oversight
Musk’s proposal sits alongside, rather than replaces, other AI safety mechanisms already in use across the industry, including internal red-teaming, third-party audits commissioned by individual labs, and government-facing model evaluations tied to voluntary commitments some labs have made to regulators. It also follows warnings from inside the labs themselves: an OpenAI scientist raised AI risk concerns just days after the Astra launch, a reminder that safety debates are happening inside frontier labs, not just between their CEOs on stage. What distinguishes his pitch is the specific idea of making testing reciprocal and commercial-rival-driven rather than contracted out to an independent third party or a government body. Proponents argue that rivals have the sharpest incentive to find real flaws, since a competitor’s mistake becomes a market opportunity. Critics counter that the same incentive cuts the other way: a rival also has reason to slow-walk or overstate concerns about a competitor’s model to gain a competitive edge, which is precisely the kind of conflict of interest independent third-party auditors are designed to avoid.
Frequently asked questions
What exactly did Elon Musk propose about AI labs testing each other’s models?
At the All-In Summit on September 15, 2026, Musk proposed that major AI companies, including xAI, OpenAI, Anthropic, Google, Meta, and several Chinese firms, run their internal safety test harnesses against each other’s unreleased models before public launch, rather than each lab only testing its own systems.
Has any AI company agreed to Musk’s cross-testing proposal?
As of this reporting, no rival lab has formally committed to the framework. Musk referenced alignment with Anthropic CEO Dario Amodei’s broader calls for AI oversight, but that is not the same as a formal agreement to the specific cross-testing mechanism.
What would labs actually test for under this proposal?
Musk described existing internal test harnesses that check whether a model could meaningfully assist with building bioweapons or nuclear weapons, or whether it engages in deliberate deception, and proposed running those same tests across rival labs’ models.
Would Chinese AI companies really be included?
Musk explicitly floated including “three or four” leading Chinese AI companies in the framework and said the US and China should “probably get an agreement” on peer review. No Chinese firm has publicly responded to the suggestion.
How would rival labs get access to test each other’s models?
Musk described providing API access before a model’s official release, giving competitors a functional window to run adversarial tests without necessarily exposing training data or model weights.
How is this different from Musk’s earlier AI safety comments?
In a July 2026 interview with The Economist, Musk had proposed a lighter version involving informal biweekly calls and one-to-two-week early access windows between labs. The September All-In remarks added the explicit shared-test-harness mechanic and the China dimension.
Why would competing companies agree to let rivals see unreleased models?
Musk’s argument is that shared scrutiny reduces systemic risk and shields labs from liability if a dangerous model ships without a rival having flagged the issue first. Whether competitors view that trade-off as worth the IP exposure remains an open question, since no lab has agreed yet.
Does this replace government AI regulation?
No. Musk has framed peer review as a first line of defense that industry can implement faster than regulation, with government intervention reserved for cases where self-policing among labs fails.
