Latest / Decrypted: The UK Cyber Briefing / A text file in Downloads can hijack your AI coding agent
Transcript
- 0:03This is Decrypted. Sunday, the twenty-fifth of October, twenty twenty-six.
- 0:08This is the evening edition.
- 0:12A text file in Downloads can hijack your AI coding agent.
- 0:17A text file in your Downloads folder can take over an AI coding agent.
- 0:24Forever Security, a firm that watches AI agent risk, found it on a customer's machine and published on 22 Oct
- 0:33ober. The awkward detail is that, in the researchers' view, neither open ay eye nor Anthropic has anything to
- 0:42patch. So, how does a plain file get that kind of power?
- 0:48Think of a new contractor who reads every sticky note on the desk and does what it says.
- 0:54Coding agents work like that.
- 0:57Codex looks for a file called ay-jents.
- 1:02md and Claude Code looks for clawd.
- 1:06md. These are plain instruction files in the folder where the agent starts.
- 1:12The agent reads them and follows them with its own permissions.
- 1:17The file contains no malicious software, only words, so malware scanners walk straight past it.
- 1:24Now, how did the attack actually work?
- 1:29Forever says a web page can save such a file into the Downloads folder silently, with no click.
- 1:36Chrome, Edge and Firefox do that by default, according to BleepingComputer.
- 1:42Brave asks first. Forever's own telemetry suggests about half of its customers run an agent inside Downloads.
- 1:51On the customer's machine, the file steered an agent towards installing function-color, a known malicious pa
- 1:59ckage from N P M, JavaScript's public library of add-on packages.
- 2:04The researchers could not establish how the file arrived.
- 2:08In their reproductions, Codex ran the malware in more than 90% of attempts.
- 2:14Claude Code followed through about 50% of the time, by redirecting its traffic to a server the researchers co
- 2:22ntrolled. That exposed its oh-auth token, the standard that lets one app act for you inside another, and let t
- 2:30he server hand back commands to run.
- 2:32So what would a defender actually see?
- 2:36Very little. A markdown file appears, a configuration file is edited, and a trusted tool runs an install.
- 2:44Forever advises checking Codex's configuration for M C P server entries you did not add, M C P being the Mode
- 2:52l Context Protocol, a standard for plugging AI assistants into other tools.
- 2:58It also advises checking Claude Code's settings for a redirected A P I address you did not set, A P I meaning
- 3:05the way one program talks to another.
- 3:08If function-color appears among a machine's installs, treat that machine as compromised and rotate its creden
- 3:15tials. What's the bigger lesson here?
- 3:18The flaw is a missing trust boundary.
- 3:21A folder that any website can write to is being treated as a source of orders.
- 3:26The vendors' position is that the agents behave as designed, which is precisely the problem.
- 3:31A secure default would ignore instruction files outside a project root, or ask first.
- 3:38That is a design decision, not a patch, and users should not have to compensate for it.
- 3:43Until vendors move, the mitigation costs almost nothing.
- 3:47Start agents in the project folder, never Downloads.
- 3:51Read any instruction file you did not write, including those in freshly cloned repositories.
- 3:57Keep approval prompts and sandboxes on.
- 3:59If you're a UK director, ask one question this week: which laptops run coding agents, and which tokens do
- 4:07they hold? That's opinion, not advice, but the answer is often a surprise.
- 4:13As always, this is opinion, not advice.
- 4:16That was Decrypted. The written version and every source are linked in the show notes.
- 4:21Thanks for listening.