Voice & Audio OpenAI Opens GPT-Live-1.1 to the API and Prices Silence at 2 Cents a Minute Coding Assistants Replit Pitted 32 Claude Agents Against NumPy Tuning: 18.9x Speedup While GPT-6.2 Swarms Collapsed at Scale Security GAO Covert Tests: All Four Top Chatbots Produced Viable Attack Plans in Under 12 Hours Once Guardrails Fell Meta Meta Puts Its 2030 Net-Zero Target Under Review as AI Data Centers Push Net Emissions Up 45 Percent Business Daniel Ek Is Backing Europa Fusion, a Stealth UK Startup Led by a Former DeepMind Researcher Betting on Pulsed Power Coding Assistants Stripe Tested 14 Coding Agents on 69 Payment Scenarios: 78% of Criteria Passed, but Only 29% of Runs Met Every Requirement Security A Transcription Firm Used by Police and Prosecutors Left Interview Transcripts and Body-Cam Footage on an Open S3 Bucket News UN Killer-Robot Talks End With a 2028 Deadline and No Binding Rules. About 100 States Wanted a Treaty; the US, Russia and Israel Blocked It Voice & Audio OpenAI Opens GPT-Live-1.1 to the API and Prices Silence at 2 Cents a Minute Coding Assistants Replit Pitted 32 Claude Agents Against NumPy Tuning: 18.9x Speedup While GPT-6.2 Swarms Collapsed at Scale Security GAO Covert Tests: All Four Top Chatbots Produced Viable Attack Plans in Under 12 Hours Once Guardrails Fell Meta Meta Puts Its 2030 Net-Zero Target Under Review as AI Data Centers Push Net Emissions Up 45 Percent Business Daniel Ek Is Backing Europa Fusion, a Stealth UK Startup Led by a Former DeepMind Researcher Betting on Pulsed Power Coding Assistants Stripe Tested 14 Coding Agents on 69 Payment Scenarios: 78% of Criteria Passed, but Only 29% of Runs Met Every Requirement Security A Transcription Firm Used by Police and Prosecutors Left Interview Transcripts and Body-Cam Footage on an Open S3 Bucket News UN Killer-Robot Talks End With a 2028 Deadline and No Binding Rules. About 100 States Wanted a Treaty; the US, Russia and Israel Blocked It

Amazon Says Attackers' Autonomous AI Agents Found a MONAI Label Zero-Day on Their Own, Then Rooted a University Server

Amazon Says Attackers' Autonomous AI Agents Found a MONAI Label Zero-Day on Their Own, Then Rooted a University Server

Amazon Threat Intelligence says attackers running autonomous AI agents on the open-source Cairn pentest framework found an unknown path-traversal zero-day in MONAI Label, CVE-2026-133968, CVSS 9.8, and reached root on a university server. The agents hit arbitrary file writes in about 11 minutes across a 48-hour September campaign. A fix was merged in October, but no patched release exists.

Amazon Threat Intelligence has reported what it describes as a first: attackers operating autonomous AI agents discovered a previously unknown vulnerability in real open-source software on their own, then used it to take root on a university server. The target was MONAI Label, an open-source medical-imaging annotation server from Project MONAI, an initiative co-founded by NVIDIA and King's College London. The flaw, tracked as CVE-2026-133968 and rated 9.8 out of 10, is a path-traversal weakness in the server's upload handling that affects versions 0.8.5 and earlier in the default configuration, which runs with authentication disabled.

Amazon was able to reconstruct the intrusion in unusual detail because the operator made a basic mistake: the control panel for their own tooling was sitting on the open internet without a password. The panel belonged to Cairn, an open-source framework that runs AI agents as automated penetration testers, and its exposed configuration revealed the agents' task settings, prompts and run history. Amazon says it traced roughly 48 hours of activity from September through the logs.

The kill chain was fast. The operator set the goal — get shell access — and picked the targets, then let the agents choose their own methods. Within about 11 minutes of starting against the MONAI Label server, the agents had mapped the application's interface and reached the point of writing files to arbitrary paths outside the designated storage folder. In a later session, they used the same primitive to plant a file that Python executes automatically at startup, then triggered a training feature that spawned a fresh Python process, loaded the planted file and connected back to the attacker with root privileges. In a subsequent session the operator handed the agents a tunnel into a second university network running the same software, which Amazon believes was compromised the same way but has not confirmed. One probing campaign against a database node ran for 16 hours.

The attribution carries heavy caveats. Amazon ties the campaign to a China-linked actor with moderate confidence, citing Simplified Chinese prompts, China Standard Time settings and targets it judged of likely interest — but it did not connect the work to any named group or establish government involvement, and the account is based on Amazon's own reconstruction, not an independent confirmation. There is no public evidence that patient scans or clinical systems were touched, and the campaign included failed operations as well as the one confirmed root result.

The patching situation is the uncomfortable part for defenders. Project MONAI maintainers validated the flaw and merged a source-code fix on October 19 — pull request 1915 forces uploads to resolve inside approved folders and rejects empty names or names beginning with a dot — and a security advisory, GHSA-qc22-rmgc-vg94, was published this week. But the advisory still lists no patched version, and the newest release on PyPI remains 0.8.5, dating to 2024. A merged fix in the repository is not a package anyone can install.

Until a fixed release ships, the practical guidance is containment: block untrusted network access to MONAI Label servers, turn on the authentication the server leaves off by default, run the service under a limited account, verify that any source build includes the October fix, and check for unexpected Python startup files and unusual training jobs — the exact path this intrusion took.

The broader lesson is about tempo. Amazon's chief information security officer CJ Moses noted that an automated agent discovering and exploiting a zero-day in about 11 minutes completes its chain faster than human analysts can triage a single alert, and that agents keep working around the clock. Research tools like MONAI Label were engineered for isolated lab networks, yet routinely end up reachable on university infrastructure alongside GPU clusters and pre-publication data. If attacker-side agents can scan never-audited codebases and return working exploits, the window between code being published and a usable exploit existing may shorten dramatically — for the large body of open-source scientific software that has never had a serious security review.

Comments (0)

Log in to join the discussion

Log In

No comments yet