Washington Asked OpenAI to Not Release GPT-5.6 — What's Next?

Study Guide

Overview

In this video, Matt Maher reacts to the US government stepping in to restrict the release of the most capable AI models from both OpenAI and Anthropic. He frames this as a genuinely big deal while deliberately trying to separate what is real from what is "overblown" hype. His central argument: the panic is happening around the weakest frontier models we will ever see, and the deeper problem is not safety alone but global competition and how (or whether) governments should control distribution.

Key Takeaways

  • The government acted twice, two different ways. It imposed an export control on Anthropic's model "by force," and made a request for OpenAI to slow the release of "5.6," wanting to vet and name the recipients.
  • "These are the dumbest frontier models we will ever see again." Maher's recurring frame: all the shutdowns, export controls, and panic are a reaction to the floor, the weakest version of a technology that only gets stronger.
  • The export-control bar is unrealistic. Demanding a model be un-jailbreakable is, he argues, like telling Microsoft it cannot ship software unless it proves the software will never have a bug. All models have always been jailbreakable.
  • The real issue is bigger than safety: global competition. If the US freezes its models, China's DeepSeek (roughly six months behind) catches up. Maher warns that in six months DeepSeek could look like today's frontier "mythos"-level model.
  • Lock-down is the wrong default. Rather than "shut it down and close our eyes," the government should ask sharper questions: define specifically what behavior is unacceptable and pre-vet models against that definition, possibly through multiple companies rather than the government itself.

How We Got Here

Project Glass Wing and "Mythos" (Anthropic)

Maher recounts that Anthropic was working on a model he refers to as mythos, which they pointed at major open-source software that runs the internet (he names OpenBSD and other long-standing projects). The model reportedly surfaced numerous zero-day vulnerabilities — security flaws sitting in the open that no one had found. Anthropic launched Project Glass Wing, inviting a set of partners to point the model at important software and patch it.

He notes the timing is suspect because Anthropic (and OpenAI) are looking to IPO, so the loud announcement could read as a marketing play. Still, he credits it as a real step: partners reportedly went on to find something like 10,000 severe/critical vulnerabilities in important "glue" software, not "little vibe-coded apps."

Fable 5 / Mythos 5 and the NSA red team

As Anthropic prepared to release, it put out Fable 5 (described as a protected version of mythos) for general public use. Information then reached the government that the model had "hacked into" some NSA network material in hours rather than days. Maher clarifies this was largely a misunderstanding or salacious framing: it was actually an NSA-run red team verifying whether the model could break into their systems, and apparently it could.

The government's reaction

Given that the breach effectively occurred under controlled testing, the government's response becomes easier to understand: it pulled the easiest lever available, an export control stating that no non-US citizen could access the model. Maher's criticism is that the "check" was never clearly defined, and even after learning it was just a red-team event, the government did not lift the control.

OpenAI and "5.6"

The government then applied pressure to OpenAI, asking it to slow the release of 5.6 and to disclose who received access. OpenAI pushed back, then agreed to roll out slowly to a small set of partners (similar to Glass Wing). Crucially, OpenAI was not placed under the same export control, so it expects to release more broadly in a few weeks.

Main Arguments

1. The control was a blunt instrument

Restricting a model to US citizens only was "the brake handle they had at the moment." That may be defensible if you genuinely fear the model will break everything, but once it became clear the NSA event was a controlled red team, the bar of "must not be jailbreakable" is, Maher argues, far too high to clear and not how LLMs work.

2. Jailbreaking is overstated

Getting past guardrails is real but costly and not reliably repeatable; it is an open, known property of how these systems work, not a switch that fully unlocks a model. So it cannot be the standard for whether a model ships.

3. The competition trap cuts both ways

Maher imagines a future where only US labs (OpenAI, Anthropic, Google) have powerful AI: all global money swarms toward the US, crushing other economies, which he argues is ultimately bad even for the US. But the opposite — freezing US models — lets China's DeepSeek catch up fast. He frames AI as the new battleground where future conflicts may be fought by "destroying economies" rather than killing people.

4. We need a defensive posture, not lockdown

His prescription: stop trying to control distribution, because it is going to happen anyway. Instead, define what must be controllable. Ask "if it does X, it's not okay to release," or "if a user could get to Y, that's not okay." Pre-vet models against concrete definitions, ideally via a number of companies rather than the government acting alone. He notes both Anthropic and OpenAI have publicly asked for exactly this kind of regulatory clarity for years.

5. A plea against hype

Maher acknowledges he is "part of the machine" on YouTube, then calls on the industry to stop amplifying every event into "loud hype noise" and to stop blaming the public for their choices. He asks for clarity over megaphone-style alarmism, especially on something this consequential for "our future" — not America's, not YouTube's, but everyone's.

Notable Quotes & Data Points

  • "These are the dumbest frontier models we will ever see again... This is the floor."
  • "It's the dumb kid walking into class and we still couldn't agree on the rules for this one."
  • Demanding an un-jailbreakable model is "akin to saying Microsoft you're not allowed to release new software unless you can prove that it will never have a bug in it."
  • Partners reportedly found roughly 10,000 severe/critical vulnerabilities after using the mythos-level model.
  • DeepSeek is described as roughly six months behind the current frontier — "just wait six months and what China DeepSeek is turning out will look like mythos."
  • "By the end of the year we have three more step changes to go through."

A Note of Balance

Maher gives the government credit for one thing: it applied controls to both Anthropic and OpenAI rather than singling out one lab. He argues an even playing field matters — businesses and the public can work within rules, but rules applied unevenly are far harder to navigate, and the OpenAI-vs-Anthropic asymmetry feels "anti-competitive at best."

Conclusion

Maher does not claim to have "golden answers." His core conclusion is that "shut it down, close our eyes, and hope" is not a viable strategy because the technology will keep advancing, competitors will keep building, and the next models will be far smarter than the ones causing today's panic. The path forward is clearer definitions of unacceptable behavior, fair and even rules, real government-industry collaboration, and far less hype.

YouTube