The attacker is now an agent — and the frameworks defenders use, turned around
An Entra application minted three accounts and disabled them five minutes later. A medical imaging tool was exploited by agents in the wild. The defence on offer: inventories, runtime checks, and a kill switch.
- An operator abused an Entra application to create three accounts and obtain tokens for an internal business API, then disabled each account roughly five minutes after creating it. m:…7886
- MONAI Label carries a path traversal flaw, CVE-2026-133968, with a CVSS score of 9.8, exploited in the wild since September 2026. m:…2741
- The fix for the MONAI Label flaw exists only on the project's main branch, merged on October 19, 2026, and has not shipped in a packaged release. m:…2741
- GPT-6.2 models meet the 'Critical' threshold for cybersecurity under OpenAI's Preparedness Framework. m:…5796
Three accounts, five minutes
In a US financial-services tenant, an operator abused an existing Microsoft Entra application to create three accounts and obtain tokens for an internal business API, according to Abnormal Security. The application held an Exchange Online role, full_access_as_app, granting access to mailboxes across the tenant. Sign-in attempts arrived six to nine seconds after each account was created, using the resource owner password credentials grant, and the application disabled each account roughly five minutes after creating it. Entra recorded the application's service principal as the initiator of each directory action, with no signed-in user behind it. It is the shape of an attack run by software, not a person at a keyboard. m:…7886
Abnormal assesses with moderate confidence that the activity is linked to ARTEX and AntiFox tooling. AntiFox is a fork of ARTEX, an agentic offensive framework. The evidence does not establish how the operator first came to control the application, and the application's original business purpose could not be verified. That honesty matters. What is documented is the pattern, and Abnormal publishes a Sigma correlation rule for detecting the create-sign-in-disable sequence. One host involved in the activity was also seen targeting two other organisations within days. m:…7886
Agents exploiting a medical tool
The same pattern, agents doing the exploiting, showed up in medical imaging software. MONAI Label, a widely used annotation tool, carries a path traversal flaw, CVE-2026-133968, with a CVSS score of 9.8, exploited in the wild since September 2026. Autonomous Cairn agents reached arbitrary file write about 11 minutes after starting API mapping. A campaign targeted university networks in East Asia. The attackers used the traversal to write a Python reverse shell into sitecustomize.py, a file Python imports automatically at interpreter start. Authentication is off by default in MONAI Label, so in many deployments the endpoint is reachable by unauthenticated clients. m:…2741
The fix exists, but only just. It was merged into the project's main branch on October 19, 2026, and at the time of writing had not shipped in a packaged release. The advisory recommends withholding the ANNOTATOR role from untrusted users. Amazon's CISO, CJ Moses, characterized the operators as suspected Chinese spies, based on script timestamps and prompts: a suspicion, not a proven attribution. The lesson for a technology executive is not about one tool. It is that the attacker is now an agent, the defender's tools can be turned around, and the window between a fix being merged and a fix being installed is where the damage happens. m:…2741
The flaw itself is mundane, which is the point. The vulnerable entry point is the image parameter of PUT /datastore/image, and the LocalDatastore.add_image function fails to verify that resolved paths remain within the datastore directory. The fix, which uses os.path.commonpath to keep destination paths inside the allowed base directory, is what has not shipped yet. m:…2741
The defence on offer: inventory, checks, and a kill switch
The labs' answer, as Straiker sets it out for security teams working with OpenAI's GPT-6.2 models, starts with knowing what you have: maintain a full inventory of AI agents, including their models, tools and permissions. The models themselves meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework — the highest level the framework assigns — which is why the advice is about operating them, not about trusting them. m:…5796
The rest of the advice is procedural. Security testing should be performed whenever models, prompts or agent configurations change, and runtime security tools should evaluate prompts and tool calls to block malicious activity. And when an agent does go wrong, there should be an Agentic Kill Switch to isolate it during the incident. None of this stops an operator abusing an Entra application or an agent exploiting a medical tool. It is a way of making sure that when it happens, someone notices fast and can cut the agent off. m:…5796
What is still unresolved
- How the operator first came to control the Entra application, and what its original business purpose was.
- When a packaged MONAI Label release carrying the fix will ship.
- Whether the operators behind the MONAI Label campaign are who Amazon's CISO suspects them to be.
- Whether runtime checks and kill switches get adopted beyond the teams that publish guidance about them.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.