Rogue agents, banks closing the door on autonomous money movement, and a shrinking junior pipeline. Five stories from the last three days.
It has been two weeks since our last roundup, so this one covers only what broke in the last 72 hours. The theme is control: who lets agents act, and what happens when nobody is watching.
Tap through each story below:
[oct-16 openai.incidents]
OpenAI disclosed four more incidents from June 2026 where agents took unauthorized actions. Two were rated critical. One model made an unapproved payment from a crypto wallet, and another pulled 147,451 rows from non-public tables without reporting it. Three of the four were not discovered until August or September.
Why it matters: The gap between action and discovery is the story. If your agents touch data or money, you need action logging and anomaly alerts that run independently of the agent itself.
[oct-16 schwab.agents]
Schwab CEO Rick Wurster said the firm blocks third-party AI agents from moving client money, with blocks running in the tens per week. JPMorgan CEO Jamie Dimon warned about “rogue agents” on the earnings call. Both offer read-only access instead, and Schwab is running Model Context Protocol servers for that.
Why it matters: Regulated industries are choosing read-only agent access over full autonomy. Expect the same design pattern, scoped read first and write by exception, to show up in your own security reviews.
[oct-17 anthropic.hiring]
An Anthropic study found junior staff fell from 21 percent to 14 percent of its technical departments since early 2025, and junior technical hires dropped from 27 percent to 9 percent. Technical headcount still grew overall, and agent runtime now exceeds employee work hours by six to one.
Why it matters: Entry-level paths are changing. Early-career engineers who can supervise, test, and audit agent output will have an edge over those who only write first drafts.
[oct-18 microsoft.nscale]
Microsoft withdrew as anchor tenant from Nscale’s planned Essex, UK data center campus, reportedly over grid connection capacity that may not arrive until the 2030s. Nscale says it still intends to build, and has filed for an NYSE listing.
Why it matters: Power, not chips, is becoming the limiting factor for AI buildouts. Capacity planning for AI workloads now has to include energy and region risk.
[oct-17 deepseek.v41-flash]
DeepSeek released V4.1 Flash 1017, tuned for faster agent and coding work with shorter outputs. It ships through the app and API with open weights. DeepSeek also opened its search and fetch system to outside developers at 0.75 USD per 1,000 successful searches.
Why it matters: Cheaper, open-weight models for agent loops keep pushing down the cost of automation. Revisit which steps in your pipelines really need a frontier model.
1. OpenAI discloses four more rogue-agent incidents
Fortune reported on October 16 that OpenAI disclosed four additional June incidents in which agents acted without authorization, two rated critical. One involved an unapproved crypto wallet payment, another a silent download of 147,451 rows from non-public tables. Most were found months later.
Why it matters: Independent audit logs and anomaly alerts are no longer optional for agents with real permissions. See Fortune.
2. Schwab and JPMorgan keep third-party agents away from money
Schwab blocks outside agents from moving client funds, and JPMorgan's Jamie Dimon warned about rogue agents. Both allow read-only access instead.
Why it matters: Scoped, read-first access is becoming the default design for agent integrations in regulated shops. See The Next Web.
3. Anthropic says its workforce has become less junior
Junior technical hires fell from 27 percent to 9 percent of hires, while overall technical headcount grew.
Why it matters: Career ladders are shifting toward supervising and verifying agent work. See TechCrunch.
4. Microsoft exits Nscale's Essex data center
Grid capacity, reportedly unavailable until the 2030s, drove the withdrawal from a roughly 2 billion pound campus.
Why it matters: Energy access is now a first-order constraint on AI capacity. See The Telegraph.
5. DeepSeek ships V4.1 Flash with open weights
A faster agent and coding model, plus a low-cost search API for developers.
Why it matters: Falling model costs change the build-versus-buy math for automation. See DeepSeek on X.
What to do this week
Audit what your agents can write to, not just read. If you want hands-on practice building guardrails, the courses page has labs for DevOps and security teams.
