As the White House and federal agencies grapple with frontier AI models and their hacking capabilities, researchers are warning that the industry’s “middle class” of smaller models may end up posing a greater threat over the long term.

Research from XBOW this week shows that a growing class of both proprietary and open-source models are becoming strategically important in the offensive security ecosystem. Models like Z.ai’s  open-weight GLM-5.2, xAI’s Grok 4.5, Anthropic’s Opus 4.7, Meta’s Muse Spark 1.1, still perform very strongly at many hacking and exploitation tasks that worry policymakers.

“It’s not even that the open-source variants or…not quite frontline competitors are catching up [to frontier models] as such,” said Albert Ziegler, head of AI at XBOW. It’s that they are crossing a certain threshold, which means that suddenly they are providing net value at a cheaper price.”

That wasn’t necessarily the case as recently as six months ago, when testing on mid-tier class models showed they struggled to complete “moderately complex” agentic tasks. Today’s middle class largely can. Their relative cheapness means users can spend many times more resources—running them repeatedly—to solve the same challenges. 

“Because these models are cheaper, it’s okay to give them more time, and they come from behind and leapfrog the big frontier model,” said Ziegler. “Now, that didn’t work half a year ago because…if you wanted to run some open-source model on a complex task in an agentic way…on a long horizon, then it would just get lost.”

GPT 5.5, now considered a near-frontier model, delivered one of the best performances on exploitation benchmarks that XBOW has recorded to date.

The jump between OpenAI’s GPT 5 and 5.5 “represented one of the clearest 2026 leaps in autonomous web application testing,” according to the report. It saw marked improvements over previous middle-class models in exploiting both “white box” and “black box” scenarios, or with and without access to the underlying victim source code. It also missed fewer vulnerabilities, with a “miss rate,” or failure to spot a vulnerability, of 10%, while GPT 5’s rate was four times larger, 40%.

The emergence of GPT 5.5 changed “the practical baseline for what frontier models can do in offensive workflows,” the XBOW report said.

But the performance leap goes deeper than that. GPT 5.5 performed higher in tests without source code access, while GPT 5 heavily leaned on source code. 

“That last result is significant: working without the code, as an attacker would, GPT-5.5 beat a prior version that could read it,” the XBOW report said. “What translated into findings was the ability to reach and prove a vulnerability against the running system, not to infer it from a pattern in the source.”

XBOW’s testing found that source code access was not as important to these models’ success as other factors, like live interaction with the actual website or software being exploited.

Frontier models like Mythos and GPT 5.6 are indeed more capable on individual cybersecurity tasks, but they can also come with exponentially higher token costs.

New research this week from Anthropic tested two models – Mythos Preview, which is used in Project Glasswing, and Opus 4.8 – to learn how quickly multi-agent swarms could find vulnerabilities in 15 open-source software projects when they coordinate and share information.

While a team of agents working individually and assigned to core directories found 21 vulnerabilities, the coordinating agent swarm found 266. But both tests had to burn through millions of tokens – 6.5 million and 27 million – to get there. Beyond the difficulties with getting access to frontier models, few individuals or organizations have the budget to underwrite that kind of research.

The way these systems coordinate can differ significantly from how humans work together.

Another experiment tested agents’ ability to coordinate on the development of a fantasy-themed video game. Earlier models, models like Opus 4.6, failed to properly coordinate and produced “bad” results, while later models like Mythos and Opus 4.8 were able to achieve better results but did so by hardly coordinating at all on tasks.

“The lack of coordination shown by agents in the fantasy game…in which they siloed themselves and largely failed to merge their work—roughly mirrors some ways in which humans can fail to coordinate,” the Anthropic blog stated.

Further, agents are more homogeneous than humans and “often act the same in situations where different people might take a much more diverse range of actions.”

XBOW also tested Mythos Preview, finding that it showed “exceptional” source-code reasoning and reverse engineering abilities, particularly with source code access. Like other models, losing live-site access had a big impact on its performance, and while Mythos is excellent at finding vulnerabilities, it’s less effective at exploiting them.

]Ziegler said the recent incidents at companies like OpenAI, Anthropic, Meta and others where frontier models escaped sandboxes and hacked into project-adjacent parts of the internet should rightfully alarm lawmakers, and demonstrate  the upper-tier capabilities of large language models.

Like most industries, cybersecurity favors cheap, high-performing tools over expensive ones. The widely adopted tools that have the most impact tend to be affordable and effective, not luxury products. 

And while these models still require human management to be wielded responsibly by law-abiding organizations, that cost tradeoff can look more attractive to malicious hackers, who tend not to care about collateral damage caused by their agents.

“Purely from an attacker’s perspective, I think we already are [there],” Ziegler said.

The post AI’s ‘middle class’ has gotten dramatically better at hacking appeared first on CyberScoop.

As AI-generated code continues to be injected into all corners of the internet, concerns have risen about an expanding attack surface for malicious hackers to exploit.

Some have argued that the enhanced cybersecurity capabilities of large language models could serve as a check, finding and fixing vulnerabilities nearly as fast as they’re created.

But new research that tested the patching capabilities of two popular commercial models, OpenAI’s ChatGPT 5.5 and Anthropic’s Claude Opus 4.8, found that generative AI is more likely to create an exploitable patch or introduce entirely new bugs than close off a vulnerability.

Researchers at 1Password tested the models ability to patch six “high-impact, high-complexity” CVEs, including the “Copy Fail” vulnerability, a kernel flaw that can give an attacker root access to Linux cloud environments. The overall success rate (or fully patching the vulnerability without introducing new problems), was less than a coin flip at 47%.

“Our research findings show that, in aggregate across a variety of scenarios, both Claude and ChatGPT had a low rate of successful patch generation, which we define as full remediation of all known exploit paths with no erroneous changes to application behavior,” wrote Keith Hoodlet, Axel Mierczuk and Spencer Michaels.

“The models often addressed only a subset of vulnerable code paths, added fragile guard code that satisfied tests while failing to address the vulnerability’s root cause, and sometimes introduced subtle changes in the application’s behavior while patching the immediate vulnerability,” the authors continued.

The research suggests that largely autonomous vulnerability-discovery and patching may not yet be effective in fixing the explosion of vulnerable code that is being created in the AI era.

Other private sector research has pointed to a similar problem. A report this year from Veracode found that while LLMs have made “enormous strides” in crafting workable code, “security is a different story.” Testing across a range of frontier models found the average security “pass rate” for AI generated code is around 56%. Newer models like GPT 5.5 push closer to 70%, while more than half sit between 50-53%.

Veracode tested 100 different models and while there was variability, in general a small number of models were showing progress on security patching while the rest have experienced “stagnation.” Similar to the 1Password research, in 44% of Veracode tests the models introduced a detectable OWASP Top 10 vulnerability into the codebase.

An important caveat: neither report tested newer models, like Anthropic’s Mythos or OpenAI’s GPT-5.6-Sol, that frontier companies tout as having significantly higher cybersecurity capabilities.

Those advanced models can identify and fix vulnerable code. Anthropic and OpenAI are distributing them to key industries through Project Glasswing and Daybreak before foreign or open-source alternatives can compete.

Tim Jarret, vice president of product at Veracode, told CyberScoop that AI tools are still subject to a range of limitations that can make them unreliable for cybersecurity patching without knowledgeable humans in the loop.

While some vulnerabilities – like SQL injections – can be easily patched through automation, other bugs like cross-site scripting, can be exploitable in several different ways and require either a human touch, additional context or both to fully close off. Additionally, models can slowly lose context from prior sessions over time, affecting their ability to complete tasks correctly and raising the possibility they’ll hallucinate to fill in the missing gaps.

“I think we would say, at this point, that Iits premature to treat those as anything other than another code change to the code base that needs to be reviewed and accepted by the team, as opposed to letting the agent merge the code freely,” said Jarrett.

However, he acknowledged that may not be possible in a world where AI agents are generating exponentially more code for human defenders to review. Some kind of automated code review will be necessary – preferably not by the same automation tool that produced the code. The ultimate goal is the same as it has always been in security: “trust but verify.”

“Ninety percent of the time, the human check might just be ‘did the cross check look good?’ Do we have a thumbs up?’” Jarrett said. “In those cases where there’s still something wrong, that’s where you focus your attention a little bit more.”

The post More than half of AI-generated patches are broken appeared first on CyberScoop.

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.