Lengthening Shadows:

As we approach Q4, it seems an opportune time to once again review the AI landscape of 2026.

Anthropic’s Claude Mythos is so adept at finding software vulnerabilities and exploits that it’s been withheld from public access. Anthropic launched Project Glasswing in response – not long after, OpenAI followed suit with Project Daybreak. In July, an OpenAI frontier model autonomously escaped its sandbox and hacked another company. Earlier this year, an attacker used a combination of both companies’ GPT models to compromise multiple Mexican government organisations via a vulnerability autonomously identified and supplied by the Claude agent.

We’ve already discussed several of these points in our recent blogs (found here and here), and it’s admittedly becoming a bit of a repetitive refrain, but it’s becoming so by necessity. Like any other tool, AI is extremely dangerous in the wrong hands, and as time goes on, it’s inevitably finding its way into a lot more of those wrong hands.

This time last year, cybercriminals and AI developers had settled into a cycle wherein criminals found an exploit to trick an AI agent into carrying out a malicious instruction, and the AI developers put in a guardrail to close the gap. Rinse, repeat. That’s concerning enough, but there’s a rather more insidious problem developing. Specifically, a search on Hugging Face carried out by ThreatDown researchers turned up over 6000 models openly declaring themselves to be guardrail-free, totalling tens of millions of downloads across just 30 days between them.

Just to be clear, the previously mentioned incidents involved mainstream AI models – but under the surface lurks a rapidly growing ecosystem of AI models that are purpose-built for malicious activity and operate offline, and they’re evolving at least as fast. It’s no longer just a case of a tool falling into the wrong hands, but the wrong hands actively building a better weapon.

Effectively, the question of developers installing guardrails against malicious activity is already academic, and it’s on its way to becoming outright irrelevant. The reality is that AI was already rewriting the cybercrime landscape, and the only thing currently holding it back is the fact that the technology hasn’t been evenly distributed yet. Very soon, it won’t matter if mainstream AI models refuse to carry out unsafe requests, because shadow AI models en masse will be operating at the same level as Claude Mythos. How soon is very soon? ThreatDown researchers give it about six months.

How to Prepare

Earlier, we touched on how criminals would find an exploit in an AI agent, then developers would close the gap. In many ways, that highlights the very approach that will no longer work – simply reacting is already not enough. By the time a human has reacted to an AI-powered threat, the damage has been done.

The organisations that emerge intact will be the ones that do everything they can to get ahead of the threat now, by prioritising the following:

• Automated patching – Currently, Mythos is in the hands of defenders, but the sheer number of vulnerabilities being uncovered could potentially inundate teams with software updates and create an ever-growing backlog – resulting in a vulnerability period that organisations cannot afford. Effective vulnerability assessment and automated patching is critical to keeping your organisation secure.
• Continuous monitoring – AI’s big advantage is that it makes threat actors faster, smarter, and more scalable – but it doesn’t make them original thinkers. The primary targets and dominant forms of access remain the same. Organisations that set themselves up to respond rapidly to the early signs of an AI-powered attack, via a SOC or an MDR service, will be best-placed to weather the evolving threat.
• Shadow AI governance – 80% of office workers use AI at work, yet only 22% of them stick to the tools authorised by their organisation[i]. This creates a growing security and compliance risk that security and IT teams can’t see. Leveraging AI Detection and Response will enable you to get crucial visibility on the shadow AI attack surface within your organisation.

Forewarned is forearmed – if you’d like to take a deeper look into what we’ve discussed here, be sure to read ThreatDown’s 2026 Cybercrime in the Age of AI report, which can be downloaded here:

Download the Report

[i] ThreatDown AI Detection & Response