Author: Merrel Briona

Shield, a technology company that builds software for IT services providers, said it has wired OpenAI’s GPT-6 Astra into its Forge platform so that AI agents can log into aging customer portals and finish routine browser work that has long tied up human technicians.

The company said the workflows are already running in production, with thousands of executions a month across its network of managed service providers. Forge is pitched as an AI operating system for IT support shops, pulling together business data and customer context to carry a ticket from start to finish.

The gap the integration targets is access. Many of the portals that technicians use daily to manage tickets, devices and software licenses predate modern APIs, so an agent has to read the screen and click through it the way a person would. Shield’s earlier attempts at so-called computer use, in which a model interprets a screen and operates it, were not dependable enough, according to Mayank Jain, the company’s Head of Technology.

“Usually, a system that doesn’t have an API is an old-school system that also doesn’t look very nice and is very hard to use, even by a human,” Jain said.

With the new model, Forge’s AI Engineer hands supported browser tasks to a computer-use agent, and technicians set the boundaries on what it may do unsupervised and what must come back for review or approval. Shield says the agents now complete those steps on their own instead of leaving a customer waiting for a person to log in.

“The only time you should have to bring in a human technician is if you need their judgment, not because you’re blocked on access,” Jain said.

The performance numbers Shield disclosed are notable for a market where reliability, not raw capability, tends to decide adoption.

Note: Figures are Shield’s own measurements of its Forge platform.

Jain said the model’s handling of interfaces was the biggest departure from prior generations, which visibly stumbled while moving through a screen. “GPT-6 Astra feels like watching a human click around,” he said.

The division of labor is not total. Phone calls, on-site visits and remote-desktop work still require people, and Shield is testing remote-desktop workflows that would let agents take on more complex actions directly on a machine’s desktop.

The timing matters for small and medium-sized businesses, which typically lean on outsourced providers for everything from a locked account to a misconfigured firewall that has knocked an office offline. Speed of response is the product those providers sell, and any step that stalls on a portal login becomes a visible delay. Shield argues that offloading routine navigation lets technicians concentrate on higher-priority incidents such as phishing attacks or critical outages.

“GPT-6 Astra allows our agents to embody what a technician does by logging into portals and legacy websites that might not have API access and performing tasks that only a human could six months ago,” Jain said. “With GPT-6 Astra, Forge is now truly an extension of our human technicians.”

The integration also signals where enterprise AI vendors are pushing next: not just drafting text or answering questions, but operating the unglamorous, undocumented software that still runs much of the business world. For IT services firms, the competitive question is shifting from whether agents can act to how much of that action can be trusted without a human watching. Shield’s answer is a permissions layer in which technicians decide in advance which tasks run unattended and which wait for sign-off.

An Arm principal software engineer who helps oversee some of the Linux kernel’s most sensitive memory-management code is pulling back from general patch review, saying the volume of machine-generated submissions has made his inbox unmanageable.

Lorenzo Stoakes, who has contributed to the kernel since 2014 and holds maintainerships covering memory mapping, madvise and reverse mapping, announced the change in a mailing-list post this week. He said he will keep those maintainer roles but will no longer review code outside the areas he owns.

“Given the current throughput in the kernel resulting from AI-generated series, I unfortunately lack the bandwidth to generally be able to offer review except in those areas I maintain,” he wrote, adding that he wanted to make his inbox “a little more manageable.”

In remarks to The Register, Stoakes said he is not leaving kernel work and has no plans to give up his maintainerships. He described the shift in his review queue as moving from occasional junk to roughly half of everything he sees. “But the AI wave is hitting [memory management] across the board, really,” he said.

The episode has put a spotlight on a widening bottleneck in open-source development: finding bugs has become cheap, while judging and repairing them has not. Microsoft’s FORGE Lab said its MDASH scanning system recently produced 59 upstream Linux kernel bug reports, one of them describing a flaw at least 21 years old. That total is modest next to the broader flow. The Linux Common Vulnerabilities and Exposures team has been averaging about 27 CVEs a day in 2026 through Nov. 2, according to ZDNET.

The same pressure is visible well beyond Linux. In its most recent Patch Tuesday release, Microsoft disclosed 726 CVEs. A large share of such reports now originates with AI tooling, leaving human reviewers as the constraint.

Greg Kroah-Hartman, who maintains the kernel’s stable releases, raised the alarm at the Open Source Summit Europe. “The onslaught of the LLM-generated patches… we don’t have enough people to review them all,” he said. “We need more reviewers. We don’t have enough. All of the maintainers are burnt out, and we have been for ten years.”

Kroah-Hartman has also argued for shrinking the number of long-term support releases the project carries, saying six is too many and that the team lacks the manpower to sustain them. He has been critical of the decade-long support model that Red Hat and SUSE attach to specific kernel versions, pointing to Debian’s approach as the better template.

There are tentative signs of relief. Linus Torvalds wrote on the Linux Kernel Mailing List that much of the new code in the Linux 7.4 rc 2 release candidate “was driven by various AI fixes,” and that in the BPF subsystem fixes are now arriving faster than new reports. He called that “a sign that we may be starting to see the tide starting to turn,” while cautioning readers not to hold their breath.

Kroah-Hartman is less sanguine. Asked whether AI-assisted fixes are catching up with AI-generated reports, he said the project is “getting closer, yes, but not quite there yet,” and declined to estimate when that point might arrive.

Daniel Steinberg, creator of the open-source cURL project and a frequent critic of low-quality machine-generated bug reports, offered a more measured view. He said the project has so far managed the flood reasonably well, though the load rises and falls depending on which contributor sends a new batch of findings in a given week.

The stakes extend past maintainer fatigue. November alone brought a serious Linux TCP networking vulnerability, a damaging OpenSSH flaw in Alpine Linux, a near-miss at Alpine’s GitLab instance and the widely unpatched PHP2Shell issue.

For companies that build products on the kernel, the review crunch matters because it lengthens the gap between a reported flaw and a shipped fix, and because maintainers under strain are more likely to miss a subtle defect buried among low-value submissions. For the Linux Foundation and its corporate backers, the pressure points toward funding dedicated review capacity rather than more scanning tools.

Note: Figures are as reported by the companies and projects involved.