Podcast · Sunday, November 8, 2026
Open-weight models now hack nearly as well as restricted ones, and criminals are already running the agents.
Today's headline
Open models now hack almost as well as the restricted ones, and criminals are already using agents to do it. A freely downloadable DeepSeek model came within eight points of OpenAI's gated cyber model, at about a seventh of the cost. Meanwhile, those same kinds of agents still fully finish only about a quarter of real coding requests. Powerful for attackers, unreliable for the rest of us.
On the clock
OpenAI will remove its GPT five point five cyber preview model from the API on February first. The notice names no successor, only the most capable cyber model your account can use. Find every call to it and plan the switch.
Security
Microsoft says a Cambodia-based group is using the ARTEX agent framework against banks and other financial firms, mostly in Southeast Asia. The agents scan the internet, then exploit two common flaws: anonymous file upload and server-side request forgery. They then install open-source remote access tools and reuse stolen cloud logins. There is no CVE here, so scanners stay silent. Patch your upload and request-forgery paths this week.
Vals AI tested agents on realistic corporate networks, starting from a public web address alone. OpenAI's restricted Daybreak model reached about eighty-one percent of objectives. The open DeepSeek model reached about seventy-three, for roughly three dollars a run instead of twenty-two. Once weights are public, no provider can restrict how they are used.
Agents
A new benchmark from Alibaba-led researchers uses real change requests from coding sessions. The best of ten models fully completed only twenty-six percent. It usually met most requirements and left one detail undone. Budget a human check on every agent-written change.
Aviator, a developer tools company, says it reviews four times more pull requests by checking five things instead of reading every line: intent, acceptance criteria, invariants, decisions and evidence. Merge time fell from two days to under four hours, by its own figures. Let machines enforce the repeated rules, and keep people on the decisions.
A researcher at the Anthropic Institute, writing in a personal capacity, tested AI models as auditors of agent logs. At a five percent false-alarm rate, one model caught nearly all reward hacking and another found about half. Use auditors to estimate how often agents misbehave, not to gate releases.
Quick ones. GitHub now lets you manage Copilot code review settings through its APIs. AWS added Kiro bots, long-running agents with their own memory. And Meta added Spaces to its Muse agents.
Business and policy
Meta's strategy chief, Dave Wehner, is leaving to become CFO of Fireworks, a platform that runs open models. Fireworks is reportedly in early talks at a thirty to forty billion dollar valuation. That figure is unconfirmed.
The New York Times found that two Chinese companies approved to buy Nvidia H200 chips have empty offices. Their executives are linked to firms sanctioned for supplying China's military. Questions now fall on how Commerce checks license terms.
In Amsterdam, police arrested up to two hundred protesters at a Microsoft data center site. A coalition says more actions are planned across Europe. Expect data center power and siting to draw scrutiny well beyond the United States.
Action items
First, patch internet-facing apps with anonymous file upload or request forgery exposure this week. Second, hunt your web servers for unexpected web shells and remote access tool downloads. Third, repoint any use of the GPT five point five cyber preview model before February first. Fourth, require a human check before you count an agent task as done. Fifth, tune any AI auditor to a false-alarm rate your reviewers can absorb.
Before you go
If you run a bank-facing or customer-facing web app, check its upload form today. Anonymous file upload is exactly where these agents knock first.