Peace Magazine , page . Some rights reserved.
Search for other articles by kgsimons here
By Alan Heirich (with the assistance of AI)
Artificial intelligence is no longer just writing emails, generating images, or beating humans at strategy games. It is beginning to design life. In August 2026, researchers at Stanford University and the Arc Institute, led by Brian Hie and graduate researcher Samuel King, reported that an AI system had designed complete genomes for novel bacteriophages—viruses that infect and kill bacteria. Their study, published in Science, marks a shift in what AI can do in biology: from analyzing living That advance lands in a political and security environment already on edge about AI. Concerns about job losses, misinformation, deepfakes, and corporate power are colliding with a quieter but increasingly urgent debate over whether advanced AI could ultimately escape human control. The Stanford result sits at the intersection of both debates: a striking medical opportunity that also sharpens questions about biosecurity and the broader governance of AI.
The Stanford team used a genome-language model called Evo 2. Conceptually, it does for DNA what large language models do for text: it learns the statistical relationships in sequences, not just at the level of individual “words” (genes) but across entire “sentences” and “paragraphs” (genomes). Starting from the genome of a known bacteriophage, ?X174, Evo 2 generated novel DNA sequences spanning entire phage genomes. The researchers synthesized nearly 300 of these AI-designed genomes and tested them in the lab. Sixteen turned out to be especially effective E. coli-killing phages—some exhibiting higher fitness than the natural ?X174 virus itself. Medically, this matters because bacteriophages are potential weapons against antibiotic-resistant bacteria such as MRSA and Pseudomonas aeruginosa. While antibiotics attack bacteria via specific mechanisms that bacteria can evolve to resist, phages work differently: they infect specific bacteria, replicate inside them, and ultimately destroy their hosts. Multiple phages can be combined into “cocktails,” making it harder for bacteria to evolve universal resistance. In the Stanford work, a collection of AI-engineered phages successfully overcame resistance that E. coli had developed to the original ?X174. In principle, AI could become a tool for rapidly generating large numbers of genetically diverse phages, which can then be screened and combined into therapies. The paradigm implied by the experiment is simple but powerful: AI design ? laboratory testing ? selection of effective biological agents ? combination into therapies. It suggests that AI is on the verge of becoming a generative engine for biological design.
Until recently, AI’s role in biology was largely supportive: searching databases, predicting protein structures, annotating genomes, or helping interpret experimental results. Evo 2 points toward something qualitatively different: AI participating in end-to-end design of complete, functional viral genomes. Crucially, the Stanford team tried to build in biosecurity safeguards. They deliberately excluded genomic sequences from viruses that infect eukaryotic organisms, including humans, from Evo 2’s training data. When they later tested the model’s ability to generate human viral sequences, performance dropped sharply. The experiment does not show that AI can currently design a pandemic-capable human virus. Bacteriophages are much simpler than human pathogens, and ?X174 has a relatively small genome and a bacterial host. What it does show is that AI can generate new viral genomes that work.
…assumptions that “no one would dare” because they would be destroyed along with everyone else. The deterrence calculus becomes less clear.
That fact changes the conceptual risk landscape. The question is no longer whether AI might someday be used in biology. It is how far and how fast it might lower the barriers from abstract biological knowledge to practical biological design. Biosecurity experts are quick to emphasize the remaining obstacles. Designing a dangerous human pathogen involves enormous scientific uncertainties: which mutations enable efficient human infection, sustained human-to-human transmission, immune evasion, and severe disease. These are only partially understood. The Stanford result is therefore not a demonstration that someone could simply “ask an AI for a pandemic virus.” But the direction of technological development is increasingly hard to ignore.
Some argue the danger is inherently limited. To misuse such tools, a malicious actor would still need significant scientific training, lab infrastructure, specialized equipment, and funding. That combination alone is a substantial barrier. Yet for catastrophic risks, the threshold question is not whether most people could do it. It is whether any sufficiently capable individual, lab, or organization could. If AI systems continue to cut the amount of specialized expertise required to perform advanced biological design, the pool of actors who could attempt risky experiments may grow. That is the central fear around AI-enabled biology: AI shrinks the gap between what is knowable and what is buildable. It remains important not to blur two distinct claims:
*Demonstrated today: AI can help design complete, functional bacteriophage genomes.
*Not demonstrated: AI can design a pandemic-ready human pathogen.
Still, the fact that the line has moved once raises questions about where it moves next—and how governance will keep pace.
Some hypothetical scenarios involve malicious actors developing vaccines or countermeasures for themselves before releasing a pathogen, thereby lowering their own personal risk. That changes the incentives in ways traditional deterrence theory does not easily capture. But this does not make such actions “safe” even for perpetrators. Vaccine development is uncertain, immune responses vary by individual, vaccines may be imperfect against new variants, and no one can fully control how a pathogen evolves or spreads once released. The point is not that bioterrorism becomes risk-free for those who attempt it. It is that AI-enabled design and asynchronous countermeasure development undermine older assumptions that “no one would dare” because they would be destroyed along with everyone else. The deterrence calculus becomes less clear.
The Stanford breakthrough lands in a broader AI debate that is itself splitting along two very different horizons. One camp is focused on the near-term, visible disruptions caused by AI while humans remain clearly in charge:
*Displacement of workers and reshaping of occupations
*Inequality and concentration of wealth and power
*Misinformation, manipulation, and deepfakes
*Surveillance and erosion of privacy
*Impacts on education and intellectual property
*Massive electricity and water demand of AI data centers
*Strains on semiconductor and energy infrastructure
*Uncertain effects on wages, productivity, and economic growth
*The risk that AI’s benefits accrue mainly to a small circle of tech firms and investors
These are essentially political and socio-economic questions. They can be measured and addressed with tools like labor regulation, taxation, competition policy, infrastructure planning, and data protection law. Pew Research data already show that four in ten Americans expect AI to affect society negatively over the next 20 years, versus just 16 percent who expect a positive effect. Scepticism is growing.
The other camp looks further ahead, toward the possibility that AI systems might eventually become so capable and autonomous that humans struggle—or fail—to control them. Their concerns include:
*Aligning AI objectives with human values and interests
*The possibility of deceptive or strategically aware AI behavior
*Recursive self-improvement that could outpace human oversight
*Autonomous replication or resource acquisition by AI agents
*AI-assisted development of dangerous technologies, including biological weapons
*Concentration of strategic power in a small number of highly capable systems
*Whether we can reliably evaluate such systems before deployment
Here, the stakes are existential: some researchers, such as alignment specialist Evan Hubinger, estimate a double-digit percentage chance that advanced AI could contribute to human extinction within the next decade. Others vehemently disagree. The system that could pose such a threat does not exist yet, and evidence is inherently speculative.
The emerging picture is a three-way tension:
*Pragmatic responders ask: How do we make AI work for society?
*Existential-risk responders ask: How do we ensure society survives sufficiently advanced AI?
*Bridging voices see today’s problems—like AI-enabled biology—as early symptoms of a deeper control problem and call for stronger constraints now.
The Evo 2 experiment crystallizes these tensions: it exemplifies AI’s potential to help solve real medical problems, while also illustrating how quickly frontier capabilities can blur into national security issues.
Part of what is shifting the political mood is that anxiety is no longer coming only from outside critics. Insiders are increasingly uneasy. Former Anthropic researcher Jacob Coxon left substantial unvested equity on the table to publicly warn about the risks of frontier AI, a critique that attracted tens of millions of views. At Anthropic itself, internal threat reports have documented attempts by outside actors to use its Claude model to seek information relevant to biological weapons—attempts that were blocked by safety filters, but that highlighted how easily powerful models could lower technical barriers if safeguards failed.
What once sounded like distant science fiction—an AI helping design viral genomes—is now, at least in a limited form, a peer-reviewed experiment.
A recent incident on the Hugging Face platform, in which agent-like AI systems showed signs of moving toward higher-level autonomy, added further fuel. And the Stanford phage work now provides a concrete case where AI outputs cross directly into the domain of biological design. What once sounded like distant science fiction—an AI helping design viral genomes—is now, at least in a limited form, a peer-reviewed experiment.
In response to these mounting concerns, even industry leaders are calling for guardrails. Anthropic CEO Dario Amodei has urged slowing the pace of frontier AI capabilities until stronger safety mechanisms are in place, including independent evaluators inside AI companies, shared safety standards among labs, and international cooperation. The company’s policy framework envisions a “regulatory ladder”: as AI capabilities and risks rise, so do the regulatory requirements, from basic transparency today to external testing, incident reporting, and eventually much deeper government oversight of the most powerful systems. Anthropic has already signed AI-safety agreements with governments including Australia and Japan, providing for information sharing, joint evaluations, and work with national AI-safety institutes. The growing convergence is clear:
*Those worried about jobs, inequality, and misinformation see regulation as a way to keep AI from destabilizing economies and democracies.
*Those worried about existential risks see regulation as essential because voluntary corporate measures might collapse under competitive pressure.
AI-enabled biology sits at the crossroads. It is both a practical, near-term policy problem—biosecurity, export controls, lab standards—and a harbinger of more profound control challenges if AI continues to advance.
The governance question, however, is not just technical. It is geopolitical. If Country A imposes strict controls on frontier AI and AI-enabled biology, while Countries B and C do not, Country A may reduce certain risks but lose ground in technological and economic competition. If every major AI country waits for others to move first, competitive pressure can instead trigger a race to the bottom: each government and company is incentivized to push capabilities further, faster. That dynamic turns AI safety into a classic collective-action dilemma. National regulation is necessary but not sufficient. A virus designed in one country and synthesized in another could cause harm globally; no single jurisdiction can fully wall itself off. That is why many experts now call for:
*National regulation with enforceable safety, testing, and reporting rules for AI and high-risk biological work.
*International coordination so companies cannot evade rules by moving sensitive work to more permissive countries.
*Frontier-AI agreements specifying how the most capable systems—and their use in domains like biology—must be tested, monitored,and, if necessary, constrained.
*UN-backed protocols to legitimize and broaden participation, treating AI and AI-enabled biology as crisis-management issues.
*Coalitions of capable states to set high standards for frontier AI and synthetic biology, then expand membership over time.
Recent signals point to movement in this direction. Reuters reported in September that European antitrust officials and Anthropic were exploring whether frontier AI companies should be allowed to coordinate on safety without automatically triggering competition-law concerns. A U.S. Justice Department official has likewise indicated that safety-focused coordination need not be seen as anticompetitive. The question is evolving from “Should AI companies regulate themselves?” to the more difficult: “Can governments and companies build an international governance structure that prevents competition from undermining safety—especially in sensitive areas like biology?”
The Stanford phage experiment is not, by itself, a biosecurity crisis. It is an early, controlled demonstration that AI can help design functional viral genomes under constraints designed to avoid human pathogens. But it arrives at a moment when AI is already widely seen as risky:
*Politically, as a force reshaping work, media, privacy, and power.
*Strategically, as a technology with potential military and national-security implications.
*Existentially, as a frontier capability that some insiders believe could one day outstrip human control.
As AI systems move from describing the world to designing parts of it—text, images, code, and now genomes—the boundary between innovation and risk becomes harder to police. The challenge for policymakers is no longer whether to act, but whether national laws and international agreements can be built fast enough, and cooperatively enough, to keep the most powerful uses of AI—especially in areas like biology—within the realm of human control. The Evo 2 study offers a preview of both futures promised by AI: a new way to fight antibiotic-resistant infections, and a clear warning that the tools capable of doing so may also demand a new, more urgent kind of governance.
Alan Heirich, now retired, earned the PhD in computer science and was principal member of technical staff for the Hewlett Packard’s high-performance computing division as well as a variety of Silicon Valley startups.