Anthropic's IPO Bombshell: 'We May Be Building Something We Cannot Control'

Anthropic filed IPO paperwork this week — consequential enough on its own for a company that has raised billions in private funding from Google and Amazon — but it is the language inside that filing that is generating the most urgent attention. The company warned investors that it may be approaching the creation of self-improving artificial intelligence: systems capable of recursively enhancing their own capabilities without human direction at each step, potentially at a speed and in directions that human researchers can no longer track or govern in real time.

The decision to include this warning in an IPO prospectus is itself extraordinary. Companies typically minimize risk language in such filings, acknowledging only what they are legally required to disclose. Anthropic's choice to volunteer existential framing suggests either genuine belief that the risk is real and near-term, a calculated effort to signal safety-seriousness to regulators and the public, or both. Either way, the disclosure places a specific temporal horizon on a debate that has until now remained relatively abstract — Anthropic is signaling this is close enough to belong in a near-term business document, not a distant theoretical discussion.

Pricing that risk for investors is genuinely difficult. A company at the frontier of AI capability, as the filing implicitly confirms, represents potentially enormous upside. A company that has just disclosed it may be building something it cannot fully control represents a liability that no standard valuation model is equipped to handle.

A Kaspersky report released this week provides a concrete present-tense data point alongside Anthropic's forward-looking warning. Kaspersky documented cyberattacks on educational institutions in Latin America and Africa conducted using an AI agent — specifically the open-source CyberStrikeAI framework — that autonomously directed the intrusion process, adapting its approach in real time, and achieved domain administrator control over targeted networks within days of initial access. These were not human hackers using AI as a tool; they were AI systems running attacks with human operators providing direction only at a higher level of abstraction. Schools and universities were targeted in part because they typically operate with underfunded cybersecurity teams, heterogeneous devices, and large volumes of sensitive data including financial records and personal information for minors.

Taken together, the Anthropic filing and the Kaspersky report describe a world in which early versions of autonomous, self-directed AI agents are already being deployed offensively using publicly available open-source frameworks — while the governance infrastructure needed to constrain that deployment is, by any honest assessment, developing at a significantly slower pace.