@DavidSacks hard to square with the timing: it landed just as that dispute was cooling, and "America First, always" from the Pentagon CIO isn't the sound of a reluctant safety regulator.
@DavidSacks and the one independent expert who saw it, Katie Moussouris, calls it Defense Oriented Prompting — defensive capability, not an exploit. Anthropic says GPT-5.5 surfaces the same known, minor vulns without any bypass.
And "this has nothing to do with the prior Pentagon fight"
@DavidSacks With respect, this is the administration's framing, not the record.
Anthropic complied — it disabled both models to follow the directive. "Dario refused to de-deploy" is simply false; that's exactly what they did.
The "jailbreak" was Amazon researchers' work.