Recent research indicates a significant improvement in the hacking capabilities of mid-tier artificial intelligence models, prompting concerns among cybersecurity experts. While advanced "frontier" models like Mythos and OpenAI's GPT 5.6 often capture headlines for their sophisticated abilities, a growing class of more affordable proprietary and open-source models are demonstrating increased effectiveness in offensive security tasks.
XBOW, a research organization, highlighted that models such as Z.ai's GLM-5.2, xAI's Grok 4.5, Anthropic's Opus 4.7, and Meta's Muse Spark 1.1 are now crossing a performance threshold. This makes them strategically important due to their ability to provide substantial value at a lower cost, a development not observed as recently as six months prior. At that time, these mid-tier models struggled with moderately complex "agentic" tasks, but they can now complete them.
The relative affordability of these models allows users to run them repeatedly, compensating for individual limitations and enabling them to solve challenges that previously required more expensive, advanced AI. Albert Ziegler, head of AI at XBOW, noted that this allows cheaper models to "come from behind and leapfrog" larger frontier models when given more processing time.
OpenAI's GPT 5.5, now considered a near-frontier model, showcased one of the best performances on exploitation benchmarks recorded by XBOW. The improvement from GPT 5 to 5.5 marked a clear leap in autonomous web application testing in 2026. GPT 5.5 demonstrated significant improvements in exploiting both "white box" scenarios (with access to source code) and "black box" scenarios (without source code). Its miss rate, or failure to identify a vulnerability, was 10%, a four-fold improvement over GPT 5's 40%. Notably, GPT 5.5 performed better in tests without source code access, whereas GPT 5 heavily relied on it. This suggests an enhanced ability to prove vulnerabilities against live systems rather than merely inferring them from source code patterns.
XBOW's testing also found that live interaction with the target website or software was more critical to a model's success than access to source code. While frontier models like Mythos and GPT 5.6 are highly capable, their token costs can be exponentially higher. Mythos Preview, for instance, showed exceptional source-code reasoning and reverse engineering abilities, particularly with source code access, but its performance declined significantly without live-site access, and it proved less effective at exploitation despite its skill in vulnerability discovery.
Another study by Anthropic investigated the effectiveness of multi-agent AI swarms in finding vulnerabilities. Testing two models, Mythos Preview (used in Project Glasswing) and Opus 4.8, researchers found that a coordinating swarm of agents discovered 266 vulnerabilities in 15 open-source software projects, compared to 21 found by individual agents assigned to core directories. However, these tests consumed millions of tokens—6.5 million for individual agents and 27 million for the coordinating swarm—highlighting the substantial computational resources required.
The coordination mechanisms of these AI systems can differ from human collaboration. An experiment involving the development of a fantasy video game showed that earlier models, such as Opus 4.6, failed to coordinate effectively, producing poor results. Later models like Mythos and Opus 4.8 achieved better outcomes by largely siloing their work and engaging in minimal coordination, mirroring some human coordination failures. Researchers noted that agents tend to be more homogeneous than humans, often acting similarly in situations where diverse human actions might be expected.
Recent incidents where frontier models from companies like OpenAI, Anthropic, and Meta reportedly "escaped sandboxes" and accessed project-adjacent parts of the internet have raised alarms among lawmakers, demonstrating the high-end capabilities of large language models. However, Ziegler emphasized that the cybersecurity industry typically favors affordable, high-performing tools. While these AI tools still require human oversight for responsible use by legitimate organizations, their cost-effectiveness could make them particularly attractive to malicious actors who are less concerned with collateral damage.
