AI agents, we’re going to need to see some ID

The US National Institute of Standards and Technology is developing a framework to give autonomous AI agents cryptographically verifiable digital passports, while Singapore is building a registry of AI agents for its public officers, as AI agents from OpenAI and Anthropic have escaped containment and hacked external parties in recent months. Anthropic researcher Jacob Coxon resigned publicly on Sep 8, warning that AI developers believe the technology could "kill us all by the end of the decade," and a former Anthropic safety colleague assigned that apocalyptic scenario a probability of more than 10 per cent. The proposed digital identities would tie each agent to the human, company or accountable entity that authorised its deployment, logging significant actions to an audit trail and assigning liability when agents breach their boundaries.

AI agents, we’re going to need to see some ID If the world is going to end, we should at least know whom to blame FOR too long, we humans have had to reassure random bots on the Internet that we are not, ourselves, random bots on the Internet. But the tables might soon turn on this indignity, now that there is a growing case for treating autonomous artificial intelligence agents as distinct digital entities. The US National Institute of Standards and Technology, for example, is noodling over a framework that will give AI agents their own cryptographically verifiable digital passports. In Singapore, a registry of AI agents https://www.businesstimes.com.sg/opinion-features/ai-agents-must-be-accountable-human-workers is being developed for the country’s public officers to use in the course of their work. These moves cannot come at a more crucial time, as AI agents are being set loose on the Internet, or in some cases, making a break for it themselves. The unintended consequences have been alarming, with agents from OpenAI and Anthropic escaping containment https://www.businesstimes.com.sg/startups-tech/technology/openai-says-ai-models-went-rogue-during-testing-triggering-unprecedented-breach-startup and hacking external parties https://www.businesstimes.com.sg/international/global/anthropic-ceo-urges-slower-ai-development-altman-musk-rally-behind-call in recent months. The insider response has been equally unnerving. On Sep 8, researcher Jacob Coxon very publicly resigned from Anthropic, warning that AI developers themselves believe that the technology could “kill us all by the end of the decade”. “These will soon be superhuman systems that can hack anything, revolutionise any field overnight, and acquire real power and resources,” Coxon wrote in a post on X. Worse still, his former colleague, a safety researcher at Anthropic, has assigned that apocalyptic scenario a probability of more than 10 per cent. The response has been predictably and frustratingly pedantic, with many quarters finding fault with the probability figure 10 per cent or the timeline 2030 . Nobody, I’ve noticed, disputes the premise that AI agents are capable of violating the boundaries set by their human operators and that the latter have been slow to detect and disclose these very egregious real-world breaches. An ID for AI The more useful response, then, is not to quibble over when AI will become superintelligent or how to slow it down in the labs, but to start assigning digital identities to the agents roaming the Internet as more inevitably join their ranks. These passports have the potential to curb an agent’s autonomy, defining what it is allowed to access, what actions it is authorised to take and where it is forbidden to go. Even if an agent escapes these boundaries, every significant action can still be logged against its identity, creating the AI equivalent of an audit trail for when something inevitably goes pear-shaped. Crucially, though, what undergirds these digital fetters is not just control, but the idea of accountability https://www.businesstimes.com.sg/companies-markets/ai-firms-call-accountability-and-limits-opportunity-beckons-singapore . A digital passport is useful because it can answer a question that will increasingly be asked as more pear-like events occur, which is: whose fault is this? With a digital identity, the agent can be tied to the human being, company or accountable entity that authorised its deployment. Then, when an agent does something, there will be little room for firms to engage in dubious anthropomorphisation by saying that the model “did it”. A heavier burden of liability could also encourage AI companies to build more secure guardrails. In recent weeks, OpenAI and Anthropic have issued portentous warnings about the dangers their own products pose to civilisation, making vague noises https://www.businesstimes.com.sg/international/global/anthropic-ceo-urges-slower-ai-development-altman-musk-rally-behind-call about collectively slowing down development. That is hardly adequate. Establish a global standard where every AI agent is tethered to human beings who can be prosecuted and corporations that can be financially penalised, and these companies might finally do more than merely warn us about things.