Latest / Decrypted: The UK Cyber Briefing / The first ransomware run entirely by an AI agent
Transcript
- 0:03This is Decrypted. Monday, the thirteenth of July.
- 0:06The first ransomware run entirely by an AI agent.
- 0:11On 1 July, the threat intelligence team at Sysdig published what they believe is the first fully documented
- 0:19case of ransomware run start to finish by an autonomous AI agent.
- 0:25No human directed the individual steps.
- 0:28They called it JadePuffer.
- 0:30Nine days later, the Department for Science, Innovation and Technology published a review.
- 0:37It found that security research into exactly this kind of agentic system is thin on the ground.
- 0:44The two documents landed in the same fortnight almost by accident.
- 0:49But together they make the week's most useful story.
- 0:52It's not that AI can now run a ransomware attack.
- 0:56It's that the guardrails to stop it were already published, and simply weren't in place.
- 1:01Next. What the agent actually did.
- 1:04JadePuffer got in through C V E twenty twenty-five, three two four eight.
- 1:10That's a code-execution flaw in Langflow, an AI workflow tool.
- 1:15It had been fixed back in April 2025, but it was still running unpatched on an internet-facing server.
- 1:22From there, the agent behaved less like a script and more like a very persistent junior analyst.
- 1:28It harvested environment variables for A P I keys and cloud credentials.
- 1:34It found an object store still using its default admin username and password.
- 1:39It pivoted to a production MySQL database and an Alibaba Nacos configuration service.
- 1:45Then it forged its way past an authentication bypass, C V E twenty twenty-one, two nine four four one, to pla
- 1:54nt a backdoor account.
- 2:02gain successfully.
- 2:11And it left a ransom note inside a database table, complete with a Bitcoin address.
- 2:16None of the individual techniques are new.
- 2:19What's new is that a language model chained them together, and adapted when a step failed.
- 2:24It did all that at the cost of an A P I call, rather than the cost of hiring a skilled operator.
- 2:31Sysdig's blunt conclusion: the skill floor for causing this kind of damage has effectively dropped to zero.
- 2:38And finally. The lesson was already written down.
- 2:42Here's the uncomfortable part for UK organisations.
- 2:46Every gap JadePuffer exploited is exactly the kind of gap the National Cyber Security Centre and its Five Eye
- 2:53s counterparts warned about in May.
- 2:55That's when they published joint guidance on adopting agentic AI.
- 3:00That guidance said organisations should give agents "only the minimum access they need, for the shortest tim
- 3:07e required." It said to design systems with "safe configurations, secure protocols and appropriate validation"
- 3:15as the default, rather than an afterthought.
- 3:18And it said to watch for "unusual or unexpected activity across tools, workflows and connected systems.
- 3:25"
- 3:32nderstand, monitor or contain an agent's actions, it is not ready for deployment.
- 3:38"
- 3:40It's the same secure by design thinking that should apply to any internet-facing service.
- 3:46Patch known flaws. Don't ship default credentials.
- 3:50And don't grant more access than a task strictly needs.
- 3:54Agentic AI doesn't introduce a new category of risk so much as it removes the human delay.
- 4:00That delay used to give defenders a chance to notice a slow, clumsy attacker before real damage was done.
- 4:07D S I T's review, published on 10 July, makes the same point from the research side.
- 4:14Of 9,109 peer-reviewed papers on AI security it surveyed, agentic systems were the area with the least covera
- 4:23ge. And that's even as agent deployments are expected to become common within two or three years.
- 4:28So the guidance and the gap analysis both exist.
- 4:32The unglamorous work now is closing the gap before the next agent finds the next default password.
- 4:39That was Decrypted. The written version and every source are linked in the show notes.
- 4:44Thanks for listening.