It was a heavy week for security updates, with Adobe and Nvidia among the vendors shipping fixes for dozens of vulnerabilities. Most of it was routine housekeeping. One batch was not, and it points to where enterprise security is heading: the software that runs autonomous AI agents.
Nvidia put out four advisories, and the headline one covers 18 vulnerabilities in NemoClaw and OpenShell, the tools it offers to run and sandbox AI agents. Two are rated critical and a dozen more are high severity, between them opening the door to code execution, privilege escalation, data tampering, information disclosure, and denial of service. There is a real irony in that: software meant to help secure AI agents turned out to carry serious flaws of its own.
One website, one hijacked agent
The most instructive of the bunch was detailed by researchers now part of Cyera, and it is unsettling in its simplicity. A single visit to a malicious webpage could let an attacker seize control of an AI agent running locally, and from there reach whatever that agent could touch, source-code repositories, cloud accounts, and more. The root cause is an old, avoidable mistake wearing new clothes. The local model service was quietly listening on every network interface while the installer assured users it was confined to the machine, letting a hostile website reach in through a technique called DNS rebinding. As one researcher put it, “local” describes where the model runs, not who can reach it.
That is the real lesson of this batch. As companies rush to deploy AI agents wired into their code, their clouds, and their data, those agents become a fat new target, and the tooling around them is still young enough that decades-old bugs are resurfacing in shiny AI packaging. Anyone running these agents locally should keep the model service bound to the loopback address, put it behind authentication, and tightly limit what the agent is allowed to reach.
The rest of the pile
Nvidia’s other fixes were more conventional but still worth applying: five flaws in its DGX Spark AI computer, a handful in its Unified Fabric Manager platform, and fresh mitigation advice for a Rowhammer-style attack against its GPUs. The week before, it had also patched code-execution bugs in its Triton inference server and its Cumulus and NVOS networking software.
Adobe’s contribution was larger in count but calmer in tone. Now publishing on a twice-monthly cadence, it released seven advisories covering dozens of vulnerabilities, including critical code-execution flaws across its Substance 3D design tools, XD, and the Campaign Classic marketing platform, plus lesser issues in Illustrator and its Content Credentials SDK. Reassuringly, Adobe says none have been exploited in the wild, and only the Campaign Classic advisory carries its higher-risk priority rating.
So which of these demands your attention today? On paper, the answer is the critical, likely-to-be-exploited items, which is the sensible way to triage any patch pile. But the AI-agent flaws deserve a second look even though nothing has been abused yet, because they mark the opening of a frontier attackers are only beginning to explore. The creative-app bugs are the familiar work of keeping software patched. The agent bugs are a preview of the fights to come.
This website uses cookies.