Interview
Could one scientist armed with AI kill a billion people?
- Richard Melange predicts that AI could accelerate the development of mirror biology from a "10 plus" year timeline to "sub 10 years," creating biological capabilities exceeding natural limits, with a specific risk that the Russian Federation may use these tools for weapons development.
- The probability of an AI-enabled viral pandemic is estimated at "one or two percent" for the current year, with expectations of increased likelihood over the "next few years" as AI capabilities advance and "mid-tier actors" (such as PhD students) experience a "strong" capability uplift by "early 25" and stronger results by "the end of 2025."
- Novices are expected to remain unable to create biological weapons due to complexity, whereas "agentic paradigms" emerging soon may allow AI to "summon from the ether" new biological capabilities without prior academic publication, enabling threat actors to perform multi-step activities previously deemed too complex.
- AI-enabled autonomy may "turn on risk chains" allowing actors to execute attacks, with a predicted shift where AI systems use a "biological switch" to stockpile pathogens for deterrence, similar to nuclear strategies, and potentially wipe out humanity or weaken responses during conflicts.
- While "AI-enabled mirror bacteria" are seen as a threat that "must never happen," defensive scenarios involving "built environment modifications" and "personalized therapeutics" aim to allow local manufacturing of countermeasures within days, potentially realized "within 20 years."
- Clinical trials are identified as the fundamental bottleneck for vaccine deployment; however, AI is expected to expedite design, moving "broad spectrum vaccines" and "multi-strain vaccines" from "preclinical pipe dream" status to "phase two" and "phase three" trials soon.
- "Data filtration" for open-weight models is predicted to become insufficient within "two years" as fine-tuning costs drop to "tens of thousands of dollars," requiring "managed access" or "trusted tester" schemes to provide national security actors with early access to frontier models.
- Governments are expected to be "unwilling to spend money" on defensive acceleration due to fiscal constraints and the non-occurrence of such threats, creating a "market failure" where private companies lack buyers for defensive R&D without government intervention.
- State-level actors like Russia, Iran, and North Korea are expected to continue upgrading biological weapons capabilities, which are reported as "not at the ceiling," while "chemical weapons" attempts enabled by AI are predicted to occur "significantly before" biological ones due to lower resource barriers.
- Defense strategies are projected to shift toward "defense dominant" subdomains, utilizing "metagenomic biosurveillance" to detect pathogen fragments and "attribution technologies" to distinguish engineered pathogens from natural ones, creating a deterrence framework based on identifying perpetrators.
- "Built environment modifications" aim to create a world where pathogens cannot exist in the air in large swathes of human habitation, potentially making the domain "defense dominant" by physically separating humans from threats and reducing the spread of all respiratory viruses.
- The outlook envisions a future where "national security actors" in the UK, US, and other nations receive "privileged access" to frontier models to accelerate defensive technologies, contrasting with the "open weight" risks that remain a "perennial problem" as compute becomes cheaper.
- "Chemical weapons" attempts remain the most common predicted AI-enabled WMD due to lower barriers, while "non-transmissible biology" attacks could still result in "millions of deaths" in concentrated urban areas, representing a deeply destabilizing risk distinct from pandemics.
- "Personalized therapeutics" are expected to be manufactured "locally" and "instantly" based on real-time surveillance, creating a decentralized defense network, while "universal vaccines" may eventually allow for the eradication of viruses similar to smallpox.
- Biosecurity is expected to become a standard component of "national security" strategy, with dedicated "biohardening" budgets and a shift from "damage control" to "defensive acceleration," driven by the need to counter "AI-enabled biological catastrophes" as a defining challenge of our time.