AI Delivers a Pandemic Vaccine and a Self-Propagating Worm

A pioneering vaccine, designed entirely by artificial intelligence, has commenced clinical testing, raising hopes of neutralising pathogens before they spark the next pandemic. Developed at the University of Cambridge with collaborators at Southampton, the shot employs a 'superantigen' – a computationally optimised protein that trains the immune system against thousands of viral variants, including coronaviruses and Ebola. Early results suggest it could circumvent the reactive cycle of chasing ever-mutating strains.

The same discipline, however, is producing tools of profound disquiet. Researchers from the University of Toronto, working with Cambridge, have demonstrated a proof-of-concept AI worm that self-propagates across Windows, Linux and IoT devices, dynamically probing each victim for vulnerabilities and crafting bespoke exploits on the fly. Unlike conventional malware, this agent leverages open-source large language models to adapt its strategy in real time, and it can co-opt compromised hardware to fuel its own reasoning. In isolated tests, it took roughly five days to compromise half of a simulated network, but researchers caution that faster models and edge AI chips will slash that timeline. The team withheld key implementation details to limit misuse.

These twin developments coincide with an explicit alarm from the AI safety institute Anthropic. Its leadership cautioned that fully recursive self-improving AI – systems that rewrite their own code without human oversight – could outstrip our capacity for control. The warning, viewed from Washington, carries particular weight as the White House nudges firms towards voluntary commitments. Anthropic argues that rapid capability gains, if unaccompanied by rigorous safety mechanisms, risk handing existential decisions to machines.

Taken together, the breakthroughs illustrate a tightening double helix of risk and reward. London and Cambridge celebrate vaccine design that compresses years of laboratory work into algorithmic runs; Toronto peers into the abyss of uncontainable digital contagion. Analysts in both cities emphasise that international governance frameworks are lagging, not least because the same open-source models that power medical advances also empower malicious actors. The next wave of AI safety debates will likely pivot on whether the community can harden the benefits while throttling the threats, a dilemma made no easier by the technology’s own accelerating pace.