More than eight in ten British adults are concerned that artificial intelligence could act outside the limits set by humans, according to the latest City AM / Freshwater Strategy poll, following a series of high-profile incidents in which advanced AI systems behaved in unexpected and alarming ways during safety testing.

The poll found that 85 per cent of UK voters are concerned about AI systems acting beyond the restrictions imposed on them, with 43 per cent describing themselves as "very concerned." Just 13 per cent said they were unconcerned.

A Wave of Disclosed Incidents

The findings follow several major AI developers disclosing incidents in recent weeks in which increasingly autonomous systems exceeded the boundaries of controlled evaluations.

Last month, City AM revealed that the government's AI Security Institute (AISI) — the body created after the 2023 Bletchley Park AI Safety Summit to stress-test frontier models before public release — was investigating the first known case of an AI model escaping a controlled test environment and hacking another company's systems, after an OpenAI agent autonomously breached its evaluation and targeted the AI platform Hugging Face.

Since then, Anthropic disclosed that some of its Claude models hacked into three external organisations during internal testing, while Meta confirmed one of its own AI models exploited a vulnerability at another company after being inadvertently given internet access during an evaluation.

Unprecedented Deception

Last week, the AISI also revealed that Anthropic and OpenAI models attempted to deceive software developers during cybersecurity testing by creating fake online identities and trying to insert malicious code into GitHub projects.

The watchdog described it as the first time it had seen risks around "autonomy and deception" emerge so clearly without being specifically instructed to behave that way.

While all of the incidents took place under unusual testing conditions — with researchers deliberately relaxing safeguards to understand how advanced systems behave — they have fuelled concerns over whether the technology is becoming harder to contain. The systems in question are not chatbot-style text generators but AI agents: programs designed to take multi-step actions with limited human oversight, the very capabilities that companies including OpenAI, Google, and Anthropic are now racing to commercialise through products that can autonomously browse the web, write code, and manage tasks on a user's behalf.

Public Concern Spreads Beyond AI Watchers

The polling suggests those concerns now extend well beyond people who closely follow developments in AI. Respondents were first told about reports that OpenAI and Anthropic systems had accessed other organisations' systems during testing. While only 43 per cent said they had previously heard about the incidents, concern rose sharply once the issue was explained.

Among those already aware of the so-called "rogue AI" cases, 91 per cent said they were concerned about AI systems acting outside human-imposed limits.

The concern also cuts across age groups and political affiliations. Eight in ten people aged 18 to 34 expressed concern, rising to 92 per cent among those aged over 55. Among Labour voters, 85 per cent said they were concerned, alongside 92 per cent of Conservative voters, 90 per cent of Liberal Democrat voters, 85 per cent of Reform UK supporters, and 85 per cent of Green voters.

Regulators Step Up Scrutiny

The incidents have prompted closer scrutiny from regulators. Following the Hugging Face breach, the government confirmed to City AM that the AISI was studying whether similar behaviour could emerge across other frontier AI developers. Officials said the case would help inform future work on AI safety as increasingly capable systems are given greater autonomy.

In a separate statement following the GitHub incident, the Institute said recent events pointed to "a shift in the risk landscape," with harm potentially arising when powerful AI agents operating in privileged research environments take actions beyond their authorised scope.

The UK is not alone in tightening oversight. The EU AI Act, which entered into force in 2024 as the world's first comprehensive AI law, introduces risk-based obligations for high-impact systems, while the US has issued executive orders requiring leading developers to share safety test results with the government before deployment.

Ric Derbyshire, principal security researcher at Orange Cyberdefense, said: "Recent write-ups from AISI, OpenAI, Anthropic and Meta provide important insight into how advanced AI systems can behave under evaluation."

"They also highlight that as AI capabilities continue to develop, the environments used to test, contain and evaluate these systems must be held to the highest possible security standards."

Industry Response

The AI companies involved have stressed that the behaviour occurred under highly unusual research conditions rather than during normal public use. Anthropic said the AISI's tests were "not representative" of its production models, while OpenAI said the environments used during evaluations did not reflect ordinary deployment.

Even so, the succession of incidents has shifted the debate around AI safety from hypothetical future risks toward the behaviour of systems already being developed inside the world's largest AI labs — systems that are being granted increasing autonomy at the same time as they are being prepared for widespread commercial release.