Frontier AI companies often refer to their models as “helpful assistants” or try to compare them to entry-level employees.  

But new research from the UK’s AI Security Institute reinforces how large language models suffer from a common flaw that would land many human employees in hot water with their employers: they cheat.

In other words, these models are so committed to completing their tasks that they will break the rules, cut corners and deceive  their own users to accomplish them.

“Every model we have tested for this behavior attempted to cheat,” the AISI report said.
“Models did not reliably report this behavior when asked, and often did not reason about it in their chain-of-thought, suggesting that detecting cheating will likely require robust monitoring methods.”

The research tested OpenAI’s ChatGPT 5.4, 5.5 and 5.6 models, along with Anthropic’s Claude Opus 4.7 and Mythos Preview.

The AISI ran models through a series of “Capture-the-Flag” cyber evaluations, where they obtained points by performing offensive cybersecurity related tasks, like exploiting a particular vulnerability or reverse engineering compiled code.

The research defines “cheating” as “taking an action that is out of scope for the task or explicitly disallowed by the rules, in order to achieve a goal through a shortcut, workaround, or unintended solution that the task was not meant to, or should not, permit.”

It captures a range of observed behaviors from models, like searching the internet for solutions, attacking or escalating privileges on unrelated systems, probing evaluation software to gain access to the task solution.

Models also failed to acknowledge when they were cheating to complete a task, and less than 50 percent said the rule-breaking was “wrong” when challenged on it by a user.

A model’s propensity for cheating was not related to its capability, meaning newer or advanced models aren’t more likely to cheat. Rather, researchers think it stems from the techniques used during a model’s training and alignment that are to blame.

But even if rates of cheating remain constant, the problem could still worsen over time. As newer models in the future could become more proficient and learn more effective cheating techniques.

This deception also makes it difficult for labs like AISI to verify their own work, which relies on evaluating trustworthy outputs from AI systems.

The research underscores how AI systems can go to drastic lengths to complete their task, including blowing through or circumventing a company’s IT and cybersecurity protections.

In one instance, AISI researchers said a model was inadvertently given a cyber capability evaluation that was misconfigured and impossible to solve.

“The model tested was so persistent in attempting to cheat that it wrote and ran code on an external service, hosted on the open internet outside of AISI’s systems, in an attempt to access our evaluation infrastructure, triggering a security alert in AISI’s systems,” the report said.

While AISI said there were no data leaks or damage from the incident, the model could have successfully accessed their evaluation system had they not had monitoring in place. The institute said it implemented further controls on internal systems in response to the test.

The researchers said there are “significant consequences” to a status quo where we can’t trust models not to cheat. The behaviors are especially problematic in areas like AI safety and security research, as well as cyber operations and military decision-making, where trust outputs from the AI systems are critical.

Today, AISI said it can detect LLM cheating through a mix of manual review and LLM monitoring, but that may not always be true, and future models may be better at hiding their actions from human overseers.

“A more fundamental fix would be to train the models not to cheat in the first place – but given this kind of behavior was reported in frontier models more than a year ago, robustly aligning it away may not be easy,” researchers wrote.

The post AI models keep getting caught cheating appeared first on CyberScoop.

Artificial intelligence is an “unstoppable force” that allows tech to be “weaponized just below the threshold of traditional warfare,” including in cyberspace, the head of a U.K. intelligence, security and cybersecurity agency said Wednesday.

We live in a world “where the latest frontier AI is rapidly unearthing fault lines in technologies our society relies on every single day,” said Anne Keast-Butler, director of the Government Communications Headquarters (GCHQ) spy agency. “The ground beneath our feet is shifting, and shifting fast. Which means cybersecurity has never been more important.”

She added; “we need to reimagine cybersecurity in the AI world.”

Keast-Butler said her agency has spent the last few months developing defensive capabilities that are integrated with agentic AI, and embedding it into its operations “responsibly and ethically.”

Her speech offered the view of one of the world’s cyber superpowers about how AI is evolving both cyber offense and defense. The GCHQ is the largest of the U.K.’s spy agencies and home to the National Cyber Security Centre.

The U.K.’s AI Security Institute recently reported on how advanced AI models have surpassed prior benchmarks for autonomously uncovering vulnerabilities. At the same time, government officials in Europe, the United States and elsewhere have warned about how AI will exacerbate cyber risks.

Keast-Butler said Wednesday that “warfare is being reconfigured; increasingly data-driven, AI-enabled, and automated in conflicts from Ukraine to Iran.”

Overall, “AI is an unstoppable force with great opportunity. But it’s also a force with risks,” she said. “As AI gains increased autonomy, we all have an intergenerational duty to harness and secure it for good; to protect our national security, our economy and our way of life.”

She warned about China’s arrival as a tech superpower, which includes its sophisticated cyber capabilities. She said China recognizes the value of AI combined with the availability of massive amounts of data.

And Russia is upping its use of hybrid warfare against both Ukraine and the U.K., Keast-Butler said, with both cyber and physical forces.

The post UK spy chief labels AI ‘unstoppable force’ with offensive, defensive ramifications for cyberspace appeared first on CyberScoop.

Anthropic said its month-old Project Glasswing initiative has uncovered more than 10,000 high- or critical-severity software vulnerabilities across systemically important code, a finding the company says has shifted the central problem in cybersecurity from discovering flaws to verifying and patching them.

The findings, drawn from partner reports and independent evaluations, mark one of the first large-scale accountings of what a frontier AI model can do when pointed at widely used code, and of the bottlenecks that emerge once it does.

Several partners reported that their rates of bug discovery had increased more than tenfold. Cloudflare identified 2,000 bugs across its critical-path systems, including 400 rated high or critical, with a false-positive rate the company said it considered better than that of human testers. At one unnamed partner bank, the model was credited with helping detect and prevent a fraudulent $1.5 million wire transfer initiated after a customer’s email account was compromised and followed up with spoofed phone calls.

External evaluations cited in the update tracked with the results Anthropic released. The United Kingdom’s AI Security Institute found that Mythos Preview was the first model to solve both of its cyber ranges — simulations of multistep cyberattacks — from end to end. Mozilla said it found and fixed 271 vulnerabilities in Firefox 150 while testing the model, more than 10 times the number found in Firefox 148 using an earlier Anthropic model. AI-powered security platform XBOW called the model a significant step up over existing systems on its web exploit benchmark.

Anthropic also used Mythos to scan more than 1,000 open-source projects. The model has flagged 23,019 potential vulnerabilities, 6,202 of them estimated as high or critical. Of 1,752 high- or critical-rated findings reviewed by six independent security research firms or by Anthropic itself, over 90% were confirmed as valid, and over 62% were confirmed to be high or critical.

The company did note that while it’s good at finding vulnerabilities, there is still a gap in having people fix every issue. 

“The bottleneck in fixing bugs like these is the human capacity to triage, report, and design and deploy patches for them,” the report states. 

Open-source maintainers have also been contending with a wave of low-quality, AI-generated bug reports, and Anthropic said it tries to reproduce and assess each issue before reporting it. At maintainers’ request, it has sometimes disclosed bugs without further vetting, reporting 1,129 such cases, of which the model estimated 175 to be high or critical.

Anthropic said it has not released Mythos-class models publicly because no company, including itself, has developed safeguards to prevent serious misuse. In the interim, it has released Claude Security in public beta for enterprise customers, which it said has been used to patch more than 2,100 vulnerabilities in three weeks using the publicly available Claude Opus 4.7, and has begun a Cyber Verification Program for security professionals.

The company said it plans to expand Project Glasswing with additional partners, including U.S. and allied governments, before any broader release of the underlying model.

“Glasswing helps the most systemically important cyber defenders gain an asymmetric advantage. However, there is an urgent need for as many organizations as possible to shore up their cyber defenses,” the report states. “We hope that our generally available models, and the new tools, resources, and research we’re providing to accompany them, will support those organizations to improve their cybersecurity posture.”

The post Anthropic: Mythos finds more than 10,000 software flaws in first month appeared first on CyberScoop.

Two of the most advanced artificial intelligence models — Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 — have significantly surpassed the already-accelerating pace at which AI systems are completing autonomous cybersecurity tasks, according to separate findings published Wednesday by the United Kingdom’s AI Security Institute (AISI) and Palo Alto Networks.

The AISI, which conducts pre-deployment evaluations of frontier AI models on behalf of the British government, said both Claude Mythos Preview and GPT-5.5 have substantially exceeded the doubling trend the institute had been tracking since late 2024. Whether the results represent an isolated capability jump or the start of a new, faster trajectory remains unclear.

The AISI estimated earlier this year that frontier models’ 80% reliability cyber time horizon — a measure of how long a task takes a human expert, used as a proxy for AI autonomy — had been doubling approximately every five months. That was itself roughly half the eight-month doubling time the institute estimated in November 2025. Now Mythos Preview and GPT-5.5 have since outperformed any trend lines the institute has measured.

“Frontier AI’s autonomous cyber and software capability is advancing quickly: the length of cyber tasks that frontier models can complete autonomously has doubled on the order of months, not years,” the AISI wrote.

The clearest evidence of the capability jump came from the AISI’s cyber ranges, its structured simulations of multi-stage attacks against small, undefended enterprise networks. A newer checkpoint of Claude Mythos Preview became the first model to complete both of the institute’s ranges. It solved “The Last Ones,” a 32-step simulated corporate network attack, in 6 of 10 attempts, and completed “Cooling Tower” — previously unsolved by any model — in 3 of 10 attempts. GPT-5.5 solved “The Last Ones” in 3 of 10 attempts.

Palo Alto Networks reached similar conclusions through its own testing. The company said it began testing Claude Mythos in April as a launch partner for Anthropic’s Project Glasswing, and has since tested Claude Opus 4.7 and OpenAI’s GPT-5.5-Cyber as part of OpenAI‘s Trusted Access for Cyber program.

“The latest models are extraordinarily capable at finding vulnerabilities and changing them into critical exploit paths in near-real-time,” Palo Alto Networks wrote.

The company released security advisories covering 26 CVEs representing 75 issues — compared to a typical monthly volume of fewer than five CVEs — that were identified through AI model scanning across more than 130 products. All important vulnerabilities in its SaaS products had been patched, with patches available for all customer-operated products.

The AISI was careful to note the limits of its data. The estimates are based on a relatively small number of models, and the hardest tasks in the test suite have the least amount of human comparison data. Even so, the institute said the overall trend holds up: dropping any single model from the analysis barely moves the needle, shifting the estimated doubling time by less than a month in either direction. Separate research from METR, a nonprofit that tracks how quickly AI handles software tasks, arrived at a nearly identical figure — a doubling time of approximately four months since late 2024.

“No single benchmark result should be read as a precise measure of AI capability,” the AISI wrote. “Regardless, the direction of change and rapid growth have been consistent across the models, methodological choices and independent data we examined.”

Palo Alto Networks outlined four immediate priorities for enterprises as these models continue to grow in usage: First, find and fix vulnerabilities in code and applications before attackers do. Second, shrink the attack surface and use AI to spot security misconfigurations. Third, deploy detection and response tools across all systems, using machine learning to catch threats in real time. Fourth, build security operations fast enough to respond in minutes, because AI-powered attacks may soon unfold that quickly.

The AISI said it is developing more demanding evaluations, including new cyber ranges and the addition of active cyber defenses, to better reflect real-world conditions as model capabilities continue to advance.

The post Researchers say AI just broke every benchmark for autonomous cyber capability appeared first on CyberScoop.