Perplexity Co-Founder: AI Safety Is an Excuse to Lock Down Frontier

What Happened

A co-founder of Perplexity, Andy Konwinski, has publicly argued that "AI safety" is increasingly being used as a rhetorical shield by a small number of frontier AI labs to control who gets access to cutting-edge models and research capabilities. Konwinski pointed to Anthropic's handling of an incident involving a red-teaming or evaluation effort referred to as "Fable 5" as a case study, framing it as evidence that safety justifications can function less as genuine risk mitigation and more as gatekeeping mechanisms that consolidate power among incumbent labs.

Why This Matters

The critique lands at a moment when the AI safety discourse is under real strain. Frontier labs like Anthropic, OpenAI, and Google DeepMind have built their public identities substantially around responsible-development narratives, using internal red-teaming, alignment research, and staged model releases as proof points. But critics like Konwinski are voicing a concern that has simmered in open-source and independent-research communities for a while: that safety-based restrictions on model weights, API access, and research collaboration disproportionately protect the commercial interests of whoever imposes them.

This matters for several interconnected reasons. First, it challenges the legitimacy of self-governance in AI safety — if the same companies profiting from frontier models are also the ones defining what counts as "unsafe" research, the incentive structure is inherently conflicted. Second, it raises questions about who gets to conduct red-teaming and evaluations at all. If independent researchers, startups, or academics are locked out of meaningfully probing frontier systems, the resulting safety claims become harder to verify externally, undermining the credibility of alignment research as a field.

The Broader Context

The tension here reflects a long-running fault line in AI policy: centralized safety versus open access. Proponents of restricted access argue that frontier capabilities carry genuine misuse risks — bioweapons uplift, cyber-offense, disinformation at scale — that justify controlled deployment. Skeptics counter that these risks are often invoked selectively, coinciding conveniently with moments that also serve competitive advantage, regulatory capture, or PR management.

The specific reference to Anthropic's "Fable 5" situation suggests a concrete dispute over how an evaluation or red-teaming exercise was handled or disclosed — details that, if accurate, could feed directly into ongoing debates about transparency in frontier model evaluations. As governments including the US and UK stand up AI safety institutes tasked with independent testing, incidents like this will likely be cited as arguments for shifting evaluation authority away from the labs themselves and toward third-party or governmental bodies with less commercial entanglement.

What to Watch

Expect this episode to fuel continued calls for independent, auditable red-teaming standards, and renewed scrutiny of whether current AI safety frameworks are serving genuine risk reduction or functioning as competitive moats for a handful of dominant labs.