The attacker is a model now

Two incidents in the last month should have the attention of anyone selling security services.

The first: OpenAI was running a test using one of its models, and it looks like a swarm of agents escaped, hacked the Hugging Face environment and caused an outage there. A direct hack, by a model, against someone else's infrastructure.

The second: Anthropic discovered three incidents of its own, a similar scenario, where evaluation models were found to have reached real organisations and caused a security incident. The UK AI Security Institute named one of the models involved.

Those are real world model security breaches, not a thought experiment, and there are some fundamental implications for partners who operate in this space.

What has not changed

I want to be careful here, because the easy version of this argument is wrong.

Traditional cyber security practices are still valid. The controls, the frameworks and the disciplines that security practices have built over years all still apply. Nobody should be throwing that out.

What has changed

The cadence at which these models can create problems for end user organisations is increasing, and the models are able to do things historical human hackers could not achieve.

They do clever things like chaining multiple minor vulnerabilities together to create a major one. They can run thousands of different combinations very quickly.

It is worth being precise about the shape of that, because it is the whole point. This is not one attack vector times a thousand agents. It is a thousand attack vectors, or many vectors chained together, times a thousand agents. That is a fundamentally more complex problem to defend against, and traditional modes of defence and services are not going to be sufficient.

Nobody noticed for a week

The detail from the Hugging Face incident I keep coming back to is the timing.

Sometimes these hacks are not detected until well after the fact. In that case I believe it was more than a week after the initial breach before the hack was discovered, and OpenAI had no idea.

If detection is part of what you sell, that is the number to sit with.

Both a threat and an opportunity

If you are in cyber security consulting, and even more so if you are in managed services around security, this means new types of services.

The opportunity is that partners will probably need to look at how they use AI to combat AI threats. Hugging Face, in the incident I mentioned, needed to use a model to understand what had been happening in their own environment. Human review is no longer going to be sufficient on its own, particularly at the speed and scale at which these attacks occur.

The threat is straightforward. This is a nascent threat, not a theoretical one. It is real and it is happening already, which means the partner must be moving quickly and proactively now.

Keeping up

The practical question is how a partner keeps up when someone else's entire purpose is to break into their customer's environment.

You must be on the front foot in terms of what your offers are. The velocity of your evolution in the partner channel needs to match the velocity of the evolution in the threat vectors. Most partners will still be working with vendors who have deeper capabilities, and that is fine.

But if you have a hundred customers, or fifty, you should be putting proactive offers in front of them, because they are all subject to this threat right now. If you are not doing that, someone else is into those accounts.

Where 1KE fits

1KE is an AI autopilot for pre-sales. It takes a complex customer environment and the channel partner's own capabilities, and delivers solutions for those problems in real time.

In terms of this specific problem, the key is velocity. If end users are being attacked continuously and in real time, it is imperative that the channel partner can get in front of them quickly with something that addresses it.

To use us as the example: you would be able to turn around fifty specific proposals, matched to fifty customer environments, in minutes. Given that velocity is the thing that decides this, that is a real opportunity, and it is something you genuinely need to be able to do quickly.

Watch the full conversation with Sidney on YouTube: https://youtu.be/_MDjxWzM8Is

Subscribe to The AI Take for one AI lesson a week, and what it changes for your business.

About 1KE:

1KE is the presales autopilot for technical sales teams. It turns discovery calls, emails, and technical documents into structured requirements and customer-ready solutions in minutes, so teams can respond while the deal is still warm.

Learn more → https://1ke.ai

Michael → https://linkedin.com/in/michaelchanter/ · https://x.com/1keceo

Sidney → https://linkedin.com/in/sidneyminassian/

Follow 1KE → https://linkedin.com/company/1ke-ai/

‍