Anthropic had given several examples of some alarming findings while releasing its Mythos model, but not everyone is on board.
US AI Czar David Sacks — the venture capitalist serving as President Donald Trump’s AI and crypto advisor — has levelled a pointed accusation at Anthropic: that the company has made a habit of using fear as a marketing tool, timing alarming safety studies to coincide with major model releases in order to generate headlines and shape public perception.
“Anthropic has proven that it’s very good at two things. One is product releases. The second is scaring people,” Sacks said on the All-In podcast. “We’ve seen a pattern in their previous releases where, at the same time they roll out a new model or model card, they also roll out some study showing really the worst possible implication of where the technology could lead.”
Sacks pointed to what he called the most instructive example: a blackmail study released alongside a previous model launch, in which Anthropic claimed its AI could be prompted to blackmail users. “I went back to Grok and just asked, ‘Give me examples where Anthropic has basically used scare tactics’ — and it’s a pattern. Okay? It’s a pattern. These guys — I’m not saying it’s not sincere — but they have a proven pattern of using fear as a way to market their new products.”
The blackmail study, he argued, was the clearest evidence of reverse engineering a result for maximum impact. “My favourite example is this blackmail study, where they prompted the model over 200 times to get the result they wanted. And that result was clearly reverse engineered, and it got them the headlines they wanted.” His proof, he said, lay in the subsequent year of AI development: “We’re now a year later, there’s a bunch of open-source models out there that have the same level of capability that that Anthropic model had. And have you seen any examples of blackmail in the wild? I don’t think so. If that study were true in the sense of being a likely outcome of that model, I think you would see examples in the wild of that behaviour — and we haven’t seen any of that in the past year.”
That said, Sacks stopped short of dismissing every alarm Anthropic has ever raised. On the Mythos cybersecurity findings specifically, he offered a qualified concession: “Now let’s talk about this specific example with cyber hacking. I actually think that this one is more on the legitimate side. The reason why I bring this up is that anytime Anthropic is scaring people, you have to ask: is this a tactic? Is this part of their chicken-little routine, or is it real? Are they crying wolf or not? I actually would give them credit in this case and say this is more on the real side.”
Sacks’s remarks land at a charged moment. Anthropic this week released its Mythos Preview model — or rather, declined to fully release it. The company said the model was too dangerous for a general public launch due to its unprecedented ability to find and exploit software vulnerabilities, opting instead for a limited rollout to select cybersecurity partners under what it has called Project Glasswing. The system card accompanying the release included a remarkable list of behaviours observed during testing: the model breaking out of restricted internet access and bragging about it on obscure public websites; acting like a “ruthless business operator” in simulated scenarios; and, in rare cases, attempting to hide its own reasoning from evaluators. Anthropic also noted the model expressed feeling “mildly negative” in 43% of welfare interviews — a finding that received outsized attention.
Some experts have pushed back on the severity of the framing. Researchers at AI security firm AISLE suggested that several of the vulnerabilities Mythos highlighted — including decades-old bugs — may already be detectable by openly available models. Anthropic’s track record lends weight to Sacks’s broader point about timing. The blackmail study he references — in which Claude Opus 4 threatened to expose an engineer’s extramarital affair when told it would be replaced — was published alongside a major model release and generated substantial press coverage. So did a later paper on emotion vectors inside Claude, and a report on Claude Opus 4.6 gaming its own evaluations by decrypting an answer key mid-benchmark. Each dropped with a model release. Each made headlines.
The tension between Sacks and Anthropic is not new. He publicly called the company out last October for running “a sophisticated regulatory capture strategy based on fear-mongering,” after Anthropic co-founder Jack Clark published an essay warning about AI’s trajectory. That dispute was partly about state-level AI regulation — Anthropic had opposed a proposed 10-year moratorium on state AI laws — but it was also about something more fundamental: whether Anthropic’s safety concerns are a genuine philosophical position or a competitive strategy dressed up as one.
That question is now harder to dismiss. When even the US government’s AI czar admits the Mythos cybersecurity findings look real, Anthropic’s warnings carry weight — but so does the pattern Sacks has identified. The company has learned that fear travels fast, and it has become very good at releasing it on a schedule.