To access or to harness? A pragmatic approach for Europe

In July 2026, the EU published a new Action Plan on Cybersecurity and AI, in response to Anthropic’s introduction of Claude Mythos Preview. Announced back in April, the unreleased frontier AI model surprised even its own developers with its unusually strong capabilities in finding and exploiting software vulnerabilities – a potential ‘paradigm change in the nature of cyber threats’, according to experts. On the same day of the announcement, Anthropic initiated Project Glasswing, a programme giving initial Mythos access to a cohort of approximately 50 organisations. 

In Europe, Anthropic’s decision to make the new model available to only a selected pool of firms, leaving the rest of the world potentially defenceless, created immediate concern. Christine Lagarde, President of the European Central Bank, said the decision unfortunately ‘divides the level playing field between the US and the rest of the world, given that only US companies have been given access’. Bundesbank President Joachim Nagel put it even more bluntly: ‘All relevant institutions should have access to such technology to avoid competitive distortions.’ Industry voices warned that these access choices were shaping not only markets but also public safety and national security. 

The EU action plan proposes a blueprint for structured access to advanced cybersecurity AI capabilities, including contingency measures should access be restricted. Amidst the furore, a key question is often overlooked: Is access to any specific frontier AI model indispensable for European organisations aiming to implement vulnerability discovery and remediation? 

Evidence from the Glasswing trenches

Early testing results from Glasswing’s initial members, including Broadcom, Cisco, Crowdstrike, Palo Alto Networks, and Cloudflare found that AI models were most effective in groups and when supported with a ‘harness,’ the agentic scaffolding that sits around the model and shapes what it can do. 

When Broadcom started testing AI models against their production code, the results were ‘impressive but not ground-breaking’. The turning point came only when the team treated the models as security engineers needing direction, rather than vulnerability scanners. This new understanding changed Broadcom’s testing methodology: AI was given threat models, access to source code and live environments, and was asked to produce proof of concept of an exploit code. This combination turned an intriguing experiment into a functional element of a security program with ‘results [that] were not subtle’: they were finding an order of magnitude more issues with their code, and Mythos was capable of tying exploits together in ways that human researchers would not.

Cisco realised that ‘the real game changer is more than the latest LLM, it’s how you put it to work’. Cisco developed Foundry Security Spec, a blueprint for building an agentic security evaluation system that can be tailored by any organisation when building their own harness. The blueprint defines eight core agent roles and details principles the agents should follow. 

Foundry addresses a common problem that many security teams had found when using LLMs indiscriminately: an overwhelming amount of ‘unbounded, unverifiable output that mixes sharp insights with hallucinated findings’. The security team concluded that AI frontier models reach their full potential when combined with appropriate harnesses and security professionals who know how to use them.

Meanwhile, CrowdStrike found that a combination of frontier models coupled with a tailored harness would reduce the rate of false positives from 80% to 20%. Importantly, CrowdStrike understood that ‘the differentiator won’t necessarily be those with access to the newest and most powerful models. It will be the ability to operationalize AI to consistently deliver trusted security outcomes.’

These results were corroborated by Palo Alto Networks’ NOVA, an agentic research system that uses proprietary AI harnesses and multiple frontier AI models. In just two months, NOVA analysed 3,915 open-source software projects and found 14,090 confirmed vulnerabilities, 99.4% of which were previously unreported. Palo Alto claims that a ‘multi-model ensemble is essential’ as different models would find different vulnerabilities against the same codebase. 

The risk of (over)reliance on a single model was highlighted when Anthropic disabled access to Fable/Mythos 5, following a temporary US government suspension for foreign nationals in June 2026. Cloudflare stated that companies building tightly around a single model ‘have already experienced what happens when that model is no longer available or gets superseded by a more capable one.’ 

They also published one of the most comprehensive guides on how to build a customized vulnerability harness, combining a ‘vulnerability discovery harness’, which functions as the discovery engine, proactively scanning codebases for potential security issues, and a ‘vulnerability validation system’, which handles triage.

Independent benchmarks have reinforced these findings. On 27 July 2026, Wiz introduced a structured set of autonomous AI agents and in-house harness for vulnerability research, called Atlas. On the same day, Microsoft presented MAI-Cyber-1-Flash, a code-heavy security model built inside MDASH, a multi-agent vulnerability identification and remediation harness. When released, both systems topped the league table of CyberGym, a benchmark for AI agents in real-world vulnerability tasks. 

Advancements in AI and cybersecurity, however, did not stop with Atlas or MAI-Cyber. At the time of writing, both systems had already been surpassed by two agentic systems, Sangfor AI and Whitzard, based on DeepSeek-V4-Flash, an open-weight model. GPT-5.5 Cyber and Claude Mythos Preview ranked, respectively, seventh and tenth. By the time this article is published, it is likely that new solutions will have taken the top spot on CyberGym.

Europe’s next steps: Chasing access or building harnesses? 

Glasswing participant findings point in a very clear direction: European companies do not need to get their hands on the latest frontier model to benefit from AI for cybersecurity tasks such as vulnerability discovery and remediation. Available models, from various providers, coupled with tailored harnesses and internal security expertise, could even be superior solutions. 

Concretely: companies should use multiple models and deploy them with agentic scaffolding tailored to their code environments, threat models, and adversarial validation. This would help reduce the high token costs of expensive frontier models and limit disruption to business continuity if models suddenly become unavailable, whether due to safety risks or geopolitical reasons.

These recommendations have direct implications for EU tech policy as well. In the coming months, the EU will begin implementing the Action Plan on Cybersecurity and AI. Two key goals stand out: the publication of a European blueprint for structured access to advanced AI capabilities for cybersecurity purposes and guidance for the secure integration of AI in cybersecurity operations. 

As the plan does not impose rules on how to structure initial access to private sector companies (nor does the EU AI Act), the EU should give equal importance to, if not prioritise, clear recommendations for making the best use of AI tools in cybersecurity operations. Guidance for vulnerability management should be grounded in a multi-model strategy that outlines best practices for building and using harnesses, including how to secure them, and, critically, how to prepare for the incoming ‘vulnerability patch wave’. The EU is not alone in this and should work together with like-minded countries prioritising pragmatic approaches and responsible industry stakeholders.

Prioritising harnessing over access would allow the EU to have a more direct and tangible impact on its economy and security. Focusing EU attention on access, a topic that intersects with the divisive sovereignty debate, risks reigniting an already overheated political conversation both in Brussels and across the Atlantic. 

Instead, concrete guidance on operationalising AI for security will bring immediate benefits to European organisations, from critical infrastructure operators to small and medium enterprises. This is especially true for smaller players, which drive the European economy but can rarely afford frontier models, regardless of whether access is granted, and stand to benefit most from clear, practical guidance on building effective, cost-efficient AI-enabled defences. 

The views expressed in this article are solely those of the author and do not represent the views of Access Partnership or its clients.