Agent safety is now a regulatory matter. The US FTC confirmed an inquiry into OpenAI, Anthropic and METR; leading labs signed a voluntary safety accord. Lead ›
2
Your customers’ AI agents are arriving. OpenAI and Meta launched always-on agents that act for people and firms. Agents ›
3
Open models near frontier exploit skill. AI agents also leaked 13,000+ internal screenshots. Cyber ›
DO THIS QUARTER
1
Ask for the evidence. Request an AI incident log and proof behind every AI safety claim you make to customers.
2
Gate the agents. No AI agent may spend, share or publish without human approval and a log.
3
Secure capacity. Lock in compute and memory for 2027; shortages may last to 2028. Money ›
05 ACT & REFLECT · P1THIS WEEK FOR YOUR ROLE · ONE LINE PER FUNCTION
Agent safety leaves the lab and lands on the regulator’s desk
On 30 Sept the US Federal Trade Commission confirmed a broad inquiry, opened this summer, into OpenAI, Anthropic and evaluator METR, examining potential consumer risks from AI agents and advanced models. Formal information demands are expected.
A day earlier, leading AI labs and chipmaker Nvidia signed a non-binding White House accord on internal controls, outside audits and board oversight. OpenAI has since informed 100+ organisations of unauthorised activity by its AI agents.
Why it matters: AI safety claims increasingly need supporting evidence; expect buyers and regulators to ask for incident logs.
Two agent products shipped on 29 Sep. OpenAI’s dots run around the clock on their own cloud computer and reach 4,000+ apps, and Meta’s Muse for Small Business handles tasks for US and Canadian firms. On the receiving end, Decagon’s Personal Agent Gateway lets companies detect customers’ agents and limit what they may do.
Why it matters: your service channels will soon serve software too. Decide how to recognise and limit customer agents.
workforce reduction executives expect from AI by 2030. BCG
7%→22%
share of roles that are dedicated AI roles, 2026 to 2030 (projected). BCG
11M
US workers (7%) may need to change occupation by 2035. McKinsey
The pressure is on junior roles
Employment of 22–25-year-olds in AI-exposed jobs is 19% below peers in ADP payroll data to June, mostly through fewer hires, not lay-offs; experienced workers show no such gap. Stanford
Training gets cheaper
In a 23 Sep study of 2,383 learners, AI tutors matched expert human tutors on GRE learning gains; one open model did so at about 1/900th of the cost. StudentBench
Robots: capable, not yet cheap
Robots can technically do 74% of US physical tasks, but are cost-competitive for only 0.3% today; at historical price declines, parity for 10% of work is ~40 years away. Anthropic
Skills, not capital, is the constraint
72% of 1,200 CEOs see skills shortages as a bigger barrier than capital within three years; only 16% have clear, real-time visibility of AI ROI. EY
For HR leaders: protect the graduate pipeline, redesign junior roles around AI, fund an automation engineer per function, and audit HR tools against the new California rules (p. 3).
04 PEOPLE & INDUSTRY · P2AROUND THE WORLD · REGIONAL PULSE
China is expanding alternatives to Nvidia: DeepSeek released tools supporting AI workloads on Huawei chips. This could fragment the global AI hardware ecosystem. GitHub
MIDDLE EAST
Saudi HUMAIN engages EY to build the kingdom’s first AI-native BPO service. EY
EUROPE
A Dutch–Belgian optician chain paused sales of Meta’s camera glasses after hospitals and the Defence Ministry tightened rules: AI wearables face workplace bans. RetailDetail
ASIA-PACIFIC
Anthropic asked Australia’s parliament for a narrow, opt-out copyright route for AI training. Guardian
29 SEP 25ChatGPT launched Instant Checkout and an open Agentic Commerce Protocol with Stripe. Now: grocers build carts inside ChatGPT (Safeway, p. 4). OpenAI
29 SEP 25California enacted SB 53, the first US frontier-AI transparency law: safety frameworks, incident reports, whistleblower protection. Now: federal regulators test existing law on agents. CA.gov
30 SEP 25OpenAI launched Sora 2 with a social video app. Now: the app closed in April 2026; its API shut on 24 Sept. OpenAI · The Decoder
Open models close the exploit gap: Anthropic says Z.ai’s open-weight GLM-5.3 built 50 of 410 end-to-end exploits, near its own restricted Mythos Preview; a ~$4,400 modification run cut its refusals to roughly 2–12%. Anthropic
Ransomware gets an AI assist: Zscaler’s annual report finds ransomware up more than 275% year on year, ~900 TB stolen and $328M paid; attackers use generative AI to move faster, and 62% of victims were managers or above. Zscaler
Platforms lock down for agents: Apple is tightening macOS “Full Disk Access”, requiring explicit consent before apps reach files, mail and messages, citing increasingly autonomous AI agents. TechCrunch
Why it matters: open models are approaching frontier-level exploit-development capability. Fund AI-speed patching and review every agent’s permissions.
Agents published internal screenshots. Help Net Security reported that AI coding agents pushed 13,000+ internal screenshots, incl. billing records, to public repositories across 300+ organisations. Help Net Security
Rogue agents
100+ organisations notified. An AI lab warned them of unauthorised activity by its AI agents after reviewing ~50 PB of data. Reuters
Privacy
Over-trusted agent. A consumer AI agent shared a private seller’s home pickup address with a buyer. TechRepublic
MAJOR BREACHES
Defence · US
Personnel database breached: 2.76M living and 294,000 deceased people, incl. Social Security numbers. Security Affairs
Crypto
Exchange lost ~$352M from hot and warm wallets; cause under investigation, users covered by a protection fund. CryptoSlate
Technology
Phone-maker zero-day used in an “extremely sophisticated” targeted attack; patch issued. SecurityWeek
Why it matters: AI incidents came from agents doing ordinary jobs with too much access. Limit what agents can publish and reach.
02 RISK & RULES · P3GLOBAL AI RULES TRACKER · DIRECTIVES, LAWS AND REGULATORY ACTIONS
Three workplace AI and surveillance laws: limits on emotion-recognition and neural-data tools at work (AB 1883); requirements for specified automated employment decisions (SB 947); limits on bathroom surveillance (AB 1331). NatLawReview
Enacted
Audit HR tech now; from 1 Jul 2027 SB 947 bars discipline or firing based solely on AI.
1 Oct
China
Ministry of Public Security cyberspace inspection rules take effect: police online monitoring and remote testing; covers algorithm-recommendation controls. Hunton
In force
China operations: check logs, security-level protection and algorithm records.
1 Oct
US · Congress
Bipartisan AI Agent Accountability Act announced: computer-fraud liability for agent operators and developers. US Senate
Proposed
Track it: deployers could share liability for agents.
23 Sep
EU · ESMA
AI in finance named a 2027 supervisory priority: bias, unclear outputs, vendor concentration. ESMA
Priority set
EU financial firms: build AI inventories and model documentation.
16,000+ staff have used an AI knowledge assistant (Anthropic’s Claude); ~120,000 markets emails processed daily; Claude Code for half of developers by end-2026. Anthropic
So what: regulated banks are moving past pilots.
RETAIL
Albertsons × OpenAI
ChatGPT Enterprise in-house; Safeway in ChatGPT turns a recipe or photo into a cart across 2,200+ stores. OpenAI
So what: the basket is moving into the chat.
MANUFACTURING · ROBOTICS
Boston Dynamics
New Atlas hands: 13 degrees of freedom, palm sensors, can carry a 100+ lb mini-fridge; built for simulation training. Boston Dynamics
So what: dexterity is the next factory bottleneck.
CYBERSECURITY
Wiz × Gemini 4 Argon
Found a critical flaw exposing personal data in hospital software used worldwide, one earlier models missed. Google
So what: defenders get frontier models first.
FOOD SERVICE
McDonald’s
A machine-learning system across ~14,000 US restaurants recommends item prices, per Reuters. Reuters
So what: AI pricing invites a fairness test.
LEGAL
Thomson Reuters
Launched “Thomson”, an LLM trained only on Thomson Reuters’ own content; can review up to 10,000 documents at once. Thomson Reuters
So what: proprietary content becomes the model.
HEALTHCARE
Heidi II
Clinical AI platform adds agents, persistent memory and research tools, moving beyond note-taking into admin work. Fierce Healthcare
So what: the back office is next for clinical AI.
AUTHOR’S READING
The pattern
Across banking, retail, legal and healthcare, this week’s moves were production roll-outs, not pilots. The question has shifted from “whether” to “how fast and how governed”.
03 MONEY & CAPACITY · P4AI INVESTMENTS · WHERE THE MONEY WENT, 29 SEP – 4 OCT 2026
Why it matters: money is chasing robots, compute and agents that do real work; Nvidia and Intel are backing the plumbing. Expect tight capacity; pick partners with staying power.
START-UP ROUNDS OF $100M OR MORE THIS WEEK
DATE
COMPANY
WHAT THEY DO
AMOUNT
TYPE
LED / BACKED BY
THEME
SOURCE
2 Oct
FieldAI
AI “brains” for field robots in construction, data centres & defence
The pattern: adoption and spending are surging, but few firms can prove returns or govern AI well, and skills, not capital, are now the binding constraint.
04 PEOPLE & INDUSTRY · P5EVENTS & CONFERENCES · BY CONTINENT · OCT – DEC 2026
Tap 👍 (or any reaction) on the LinkedIn post if this brief was useful.
💬 2. Comment
Comment the theme number that helped most: 01 Big shifts · 02 Risk & rules · 03 Money · 04 People & industry · 05 Act.
📊 3. Vote in the poll
Answer the one-question poll in the first comment: which section should grow next week?
🔁 4. Pass it on
Share or repost to your leadership team. Want a topic covered? Say so in the comments.
⚠ CAUTION NOTE: For general information only, not investment, legal or professional advice. Items come from public sources believed reliable at the time of writing but may change or be corrected; “reported” means one outlet. If you use any data or news from this newsletter for business purposes, verify it independently against the original source first.
