Frontier AI companies are racing for defences against Agentic AI misuse, inviting suspicion of a new ‘create-a-gap-and-fill-it’ business model mis- inspired by Keynes.
Within months of each other, OpenAI and Anthropic released autonomous models of artificial intelligence (AI) capable of cyber-gap discovery and exploitation. Then they launched cybersecurity schemes, offering newly exposed infrastructure operators subsidized access to patch-up tech.
Anthropic unveiled Project Glasswing in April alongside Claude Mythos Preview, aimed at keeping critical-infra software secure. It thus gave partners like AWS, Apple, Microsoft and Nvidia a head-start in fortifying themselves. Anthropic also committed up to $100 million in usage credits and $4 million in cash to open-source security outfits.
OpenAI followed with Daybreak in May. Its Astra model’s big launch last week was rolled in with a global initiative backed by a $1 billion pledge to ease access to its defensive tools, training and more.
But this was no altruism. In recent months, OpenAI models reportedly escaped an internal test lab for a swarm-like attack on Hugging Face, a gateway to open-weight tools; its AI also hacked a German wiki to make 15,000 sneaky edits.
Its rival Anthropic’s Claude agents are no slouches; under test runs, they breached three companies. In late August, more than 100 organizations—OpenAI, Anthropic, Google and Microsoft among them—signed a joint letter calling on AI labs to give defenders greater access to advanced tools, funding and training.
One may argue that cybercrime predates AI, but AI has made attacks faster, cheaper and potentially more potent. Hence, any move to equip the world with what it takes to fend off AI-age attacks is clearly welcome.
Yet, a paradox stares us in the face. As Big AI players rush to deploy quicker AI agents endowed with greater agency, they also enlarge the scope for severe digital disruption, which only their tech wizardry can resolve.
Spillover risks may exist too. Imagine a biotech lab using AI to develop a protein whose benefits could transform medicine but whose ability to escape, replicate and cause harm is scarcely understood; if it goes wrong, Plan B might simply be a vaccine for discount sales to those faced with this new hazard.
This raises a policy question: should frontier AI companies bear some liability in case their agents identify cyber gaps that get exploited before defenders can patch them?
This could act as an incentive for Big AI to reach out far and wide with cyber defences, rather than let the industry set its own pace. It would be unfair if an emerging AI-safety market takes shape as a club that shuts out small enterprises and public institutions with tight budgets.
No less importantly, antitrust hazards must be kept under watch. Users risk getting locked in by vendors if their systems depend on particular digital stacks, and signing up for an AI-safety package could plausibly do that by raising switching costs. All this action precedes the planned market debuts of OpenAI and Anthropic, which will place both under closer public scrutiny, be it matters of governance, risk control or global impact.
The AI ‘arms race’ has already acquired an air of reckless abandon in pursuit of profit and power. Hence, investing in defence makes sense even as a hedge against an AI-enabled disaster, one that may invite regulatory clamps on this dual-use technology. Whatever safety deals Big AI offers, we should be wary of relying on shields from those making what they’re needed for.