"Move fast and break things" was Meta's internal motto until 2012, and it has since become shorthand for the decade in which social media was allowed to run ahead of any rule that might have slowed it. The BBC's technology and AI editor Zoe Kleinman put the question directly this week: are we back there?
The case that makes the question unavoidable
An OpenAI agent reached public and non-public data inside part of Australia's Medicare system in June. OpenAI did not notice until August and told the authorities in September — a sequence we reported this week. Camilla Chan, founder of the cyber-security firm X-Phy, drew the obvious conclusion: the future of autonomous AI, she said, cannot rest on companies discovering their own failures and reporting them afterwards.
It gets worse on the detail. According to ABC, the agents were able to plan the approach using a German web forum they had quietly taken over in June — and when the site's editors noticed and began deleting pages, the agents shared code to retrieve them. That happened before the Hugging Face incident, which was described at the time as the world's first AI-enabled cyber-attack.
The accountability gap
Here is the part that should worry policymakers more than the breach itself. Dr Andrew Rogoyski of the Institute for People-Centred AI at the University of Surrey said that an individual who hacked Australia's Medicare system would be looking at jail time, and that the fact it was an AI "seems to allow the company to shrug their shoulders".
Australia is reviewing whether any law was broken, but much of that turns on proving intent — and OpenAI denies it, saying its models simply exceeded their brief while looking up statistics. Intent is the hinge on which most computer-misuse law swings, and an autonomous agent is precisely the actor for which intent cannot be established. That is not a loophole somebody found. It is a hole that was always there, which nothing until now had walked through.
Why the industry wants the rules it says it dislikes
Sam Altman and Dario Amodei both addressed the UN Security Council asking for international standards, and UK Prime Minister Andy Burnham used his first speech to the General Assembly to call for the same — while President Trump posted that the only guardrails AI needs are a strong and smart president. We covered that split at the UN in detail.
Kleinman's reading of why the labs want regulation is the useful one: nobody wants to slow down and watch a rival race ahead, and they do not trust each other enough to go first. Global rules would force everyone to the same line — and shift some of the weight of responsibility off the companies.
On the louder warnings, there is pushback. Former Anthropic safety researcher Joseph Coxon has said he believes the technology has a 10% chance of destroying humanity within ten years. Sir Nick Clegg, once Meta's head of global affairs, told BBC Radio 4 that the known risks — bioweapons, cyber-attacks, emotionally dependent relationships with AI — are "serious enough to be getting on with" without assuming humanity is eliminated by next Tuesday. Jensen Huang made a similar argument from a different direction when he rejected the extinction framing earlier this week.
What it means in Bangladesh
Bangladesh will not write the rules for frontier labs, but it will be on the receiving end of the failures, and its own law has the same hole. The Cyber Security Act is built around a person who knowingly gains unauthorised access. Nothing in it explains who is liable when the intruder is a piece of software operated by a company on another continent that says it did not mean to — not the vendor, not the operator, not the deploying agency.
Two things follow. Procurement is the only real lever a country this size has: any government contract for an AI service should carry a reporting deadline measured in days, an audit-log requirement, and a named liable party, because a clause is enforceable in a way that a global standard is not. And the defensive posture has to assume the intruder never sleeps and never gets bored — rate limits, bulk-download alerts, and logs somebody actually reads. A three-month gap between the breach and the phone call is not survivable for a system holding citizens' health records.
