Zuckerberg said he thinks some AI labs are starting to build and train more advanced models and then not release them.

Judge the claim's substance, as of 2026-09-15: some AI labs were starting to build and train more advanced models and then not release them (publicly or generally). It is hedged ('I think', 'starting to'), so a few clear cases are enough. True: credible evidence (lab announcements, system cards, or reliable reporting) shows at least one, and preferably several, AI labs had trained models more capable than their public releases and chose to withhold them or limit access to a small set of partners, instead of releasing them generally. Mostly true: the evidence shows withheld or restricted-access advanced models, but mainly from a single lab, or access was limited instead of fully withheld. Mixed: the cases mostly look like normal pre-release delays, not decisions to withhold, with only ambiguous examples of models being held back. Mostly false: the only evidence is speculation or normal staged rollouts, with no credible case of a lab deliberately not releasing a more advanced model. False: credible evidence shows labs were releasing their most advanced models promptly, with no withholding.

Set before the research began.

1. E1 Claude Fable 5 and Claude Mythos 5 Anthropic · 9 June 2026

   “Claude Mythos 5 is restricted to Glasswing partners (with cyber safeguards lifted) and soon to select biology researchers (with biology and chemistry safeguards lifted) only, until our broader trusted access program is available.”

2. E2 Claude Fable 5 and Claude Mythos 5 Anthropic · 9 June 2026

   “In April we began Project Glasswing, releasing the first Mythos-class model (Claude Mythos Preview) to only a limited group of cyber defenders and critical software infrastructure providers.”

3. E3 OpenAI to limit access to Astra model's advanced cyber features due to hacking concerns Fortune · 1 September 2026

   “But only a handful of partners will get access to its most advanced cybersecurity capabilities as OpenAI works to balance helping companies prevent cyberattacks while not empowering attackers at the same time”

4. E4 Amodei says Anthropic's safety delay may have cost it an AI lead The New Stack · 21 October 2026

   “Anthropic first released Claude Mythos Preview in April through Project Glasswing, giving selected partners, including Amazon, Apple, Microsoft, CrowdStrike, and the Linux Foundation, access to the model’s vulnerability-discovery capabilities for defensive work while keeping it out of reach of ordinary developers.”

Anthropic's June 2026 announcement says its Claude Mythos 5 model was restricted to Glasswing partners, and soon to select biology researchers, until a broader trusted access program is available [E1]. It also says Claude Mythos Preview went in April only to a limited group of cyber defenders (E1, E2). The New Stack independently reports that this model was kept out of reach of ordinary developers (E4). Fortune reports that OpenAI would give only a handful of partners access to its Astra model's most advanced cybersecurity capabilities (E3). That is a second, less clear-cut case, because it restricts certain capabilities rather than withholding a whole model [E3]. The evidence shows advanced models with restricted access, but the clearest cases come mainly from one lab, and access was limited to partners rather than fully withheld. That matches the criteria for mostly true.