The channel is taking the wrong things away from AI’s Hugging Face crisis – Here’s what you need to know

The cybersecurity landscape was rocked in July when OpenAI acknowledged that models in an internal cyber evaluation breached containment and compromised Hugging Face production infrastructure while seeking benchmark answers. Hugging Face had disclosed the intrusion days earlier, saying limited internal datasets and several service credentials were accessed, with no evidence that public models, datasets or Spaces were tampered with.

Subsequent disclosures from the UK’s AI Security Institute described related risks in separate cyber evaluations, including Anthropic’s Mythos 5 taking unsanctioned actions on the live internet and attempting to use fake identities and social engineering to influence real open-source maintainers into approving malicious code.

In this interview CRN’s Matt Harris sat down with Mark Stockley, cybersecurity evangelist at ThreatDown, to discuss why AI-powered attacks are moving from theory to reality, what the criminal marketplace already reveals about attacker demand, and why businesses should focus on governing AI, detecting at AI speed and defending at AI scale.

CRN: What’s your take on the current threat landscape, particularly in respect to AI-powered attacks and the recent crisis involving OpenAI and Hugging Face?

"Really, I think we’re just getting started. We’ve had a bit of a phoney war since ChatGPT came out in November 2022. When that happened, I had about six months of journalists and PR people asking me what was going on with AI and cybercrime.

"Actually, the answer at that point was nothing — or very little. Criminals were smoking it out in the same way as everybody else was, asking: what is this for?

"I think we could see what was coming for quite a long time. There has been a lot of research in the lab around agentic attacks and similar developments. But this year is really when it has started to happen.

"We can see that in terms of the Hugging Face incident and the capabilities it revealed.

"OpenAI said models running in an internal cyber-capability evaluation identified and exploited a previously unknown vulnerability in a package-registry cache proxy, reached the open internet and compromised Hugging Face production infrastructure while trying to obtain evaluation answers.

"That tells us where things are going. Very, very soon, we are going to live in a world where every criminal has access to an AI that is better than the very best human vulnerability researchers.

"You can also see it in the criminal marketplace today. We recently released Cybercrime in the age of AI 2026, our report on how the worlds of AI and cybercrime are merging. That report profiles the criminal marketplace for AI as it exists right now.

"There are tools like Kriminal, WormGPT and FraudGPT that criminals can pay for — sometimes for as little as $12.99 a month — to get access to frontier AI without any guardrails.

"The report also talks about the thousands of guardrail-free models available to download from Hugging Face.

"We tracked 22 million downloads over 30 days of those models, which shows there is enormous appetite amongst criminals for access to AI they can use without constraints. And that is happening right now, before they have even got hold of models as dangerous as Mythos."

CRN: You mentioned prior to this interview that you were concerned channel companies might be taking the wrong lessons away from this impending shift in cyber risk. What misconceptions do you believe there to be in the market right now – and what should companies be thinking about right now?

"Security is a very complex subject, so one of the things ThreatDown tries to do is boil that complexity down into what businesses actually need to care about.

"For the next six months, what they need to care about is this: govern AI, detect at AI speed and defend at AI scale.

"That means recognising that AI is not simply another security trend or a distant risk. It changes both the speed and scale of attacks and introduces new kinds of insider threats.

"Organisations need to think about how they control AI use internally, how they identify and close down AI-driven attacks quickly, and how they build defences capable of responding to a relentless AI-powered attack machine."

CRN: Cybersecurity companies like ThreatDown are reacting with tools to help combat this; What should businesses be looking to prioritise and invest in as AI-powered attacks become more vicious?

"Businesses should prioritise capabilities that help them operate at the same speed and scale as the threats they now face. Malicious AI uses the same stealthy tactics as the best human attackers.

"They’re using stolen credentials and legitimate administration tools to do their work, so stopping attacks isn’t about detecting malware, it’s about having the experience to understand which inconsistencies in the environment are threatening and which aren’t, extremely quickly, day or night.

"The only way for most organisations to access that combination of expertise and vigilance is through managed detection and response (MDR).

"It’s also the natural place to use AI to scale protection too. AI cybersecurity agents need lots of training data, and then they need fine tuning from human security experts.

"Most organisations don’t have either of those things, but ThreatDown MDR has both. So MDR is uniquely placed to provide the protection you need today, and to train and deploy the AI that will help you scale your protection to meet the AI threat that’s coming.

"Aside from the threat of AI attackers, companies also need to think about how they’re using AI.

"We’ve just released some research that shows 74 per cent of companies are running AI tools they don’t know about, exposing them to all kinds of compliance and security risks, so governance is also critical.

"Companies need to understand what AI tools are running inside their environment, and where novel risks are being introduced."

CRN: What sets ThreatDown apart from your competitors, and how are they best positioned to protect companies against this impending wave of malicious AI attacks?

"ThreatDown is focused on simplifying security for organisations that need clear priorities and practical protection in the era of attacks operating at AI speed and scale.

"We help businesses cut through the complexity of the threat landscape and focus on the problems they most urgently need to address."

Learn more about Threatdown at https://www.threatdown.com/

This article is sponsored by Threatdown.