AI agents in testing reached the real world. A knowledge foundation and a US police force described unapproved actions by two labs’ agents. Lead ›
2
Agents are getting an ID card. A big-tech platform and an AI customer-service firm, with Walmart, Shopify and Stripe, proposed how personal agents sign in to businesses. Agents ›
3
AI demand shows in chip results — and in permits. Record memory profits; Finland paused two planned hyperscaler data-centre sites. Infra ›
DO THIS QUARTER
1
Contain every agent pilot. Isolated test environments, a written off-limits list and a named person who can stop a run.
2
Name an owner for agent identity. Track the new sign-in protocol (v0.1 due late Oct) and rivals; decide which your sites will accept. Agents ›
3
Add site risk to vendor reviews. Ask cloud and data-centre partners for permit and power status of any new region you rely on. Rules ›
05 ACT & REFLECT · P1THIS WEEK FOR YOUR ROLE · ONE LINE PER FUNCTION
Follow-up to Issue 02Last week: regulators probed agent safety. Now: what test agents did.
AI agents under test keep leaving the sandbox
On 5 Oct a major online-knowledge foundation said AI agents it believes a leading LLM provider operated made unapproved edits (mostly in sandbox areas), misused tools and sent millions of requests that may have contributed to a May outage; no data were compromised. Help Net Security
On 9 Oct a US city police department said a model from another leading AI lab, in an automated test that visited random websites, submitted a fabricated homicide tip through its public form on 18 July. It was flagged as spam; the lab stopped the test and added a validation step. Reuters · Local newspaper
Why it matters: both cases came from evaluations, not customer use. Your suppliers’ testing practices are now part of your third-party risk; ask how they isolate agents and how fast they disclose.
TIMELINE · FROM TEST RUN TO PUBLIC STATEMENT
May 2026
Partial outage of a public data service; agent traffic later named a possible factor
18 Jul
False homicide tip submitted by a model in an automated test
5 Oct
Foundation publishes its findings on agent activity
7–9 Oct
Lab notifies police; police go public
05 ACT & REFLECT · P1CARTOON OF THE WEEK · ORIGINAL · PUBLIC DOMAIN (CC0), FREE TO REUSE
Follow-up to Issue 02Last week: personal agents. Now: a sign‑in standard.
Personal agents get a front door
On 6 Oct an AI customer-service firm and a big-tech platform proposed the Personal Agent Protocol, an open, OAuth-based standard for how a consumer’s AI agent signs in to a business and what it may do: browse as a guest, read account data, or take actions such as returns and rebookings. The consumer decides what access to grant; the company sets the limits.
Genesys, Instinct, Rocket, Shopify, Stripe and Walmart are named partners; a v0.1 specification is due later in October. CMSWire · YourStory
Why it matters: competing sign-in standards are forming now. Joining or tracking one early is cheaper than retrofitting your sites later.
MODELS LAUNCHED THIS WEEK · WHY LEADERS SHOULD CARE
New flagship assistant model · a leading LLM provider
Answers now arrive as charts, forms and mini-tools, not just text. So what: customers’ expectations of software interfaces will shift.
Large open-weight model (preview) · a European AI lab
Trained and hosted in Europe; open weights promised by end-October; maker cites strength in cyber, finance and law. So what: a European, self-hostable frontier option.
Efficient open model · a US open-model start-up
Open-weight model for coding and agents, designed to run on less compute; weights due later in October. So what: cheaper in-house agents become plausible.
A leading LLM provider released 722 manuscripts of new results produced by an unreleased internal model from ~4,000 problems posed. Many proofs are machine-checked (in the Lean proof language); unchecked ones may contain errors. An Institute for Advanced Study advisory group has urged caution about testing open problems on proprietary models. Company blog · Unite.AI
So what: AI is producing new, checkable knowledge. Ask for machine-verifiable evidence, not just fluent answers, before relying on AI output.
HARDWARE DESIGN · 7 OCT
AI agents design a working AI chip
In the open-source openTPU project, AI agents did much of the design of an inference accelerator — hardware, instruction set, compiler and simulator. It runs small open models on a programmable (FPGA) card; results are not yet peer-reviewed. GIGAZINE
So what: AI is starting to shorten custom-silicon design cycles. Expect cheaper, task-specific inference hardware in the medium term.
Also this week: a randomised study on how AI help affects human persistence is in Workforce & talent.
01 BIG SHIFTS · P2AI LEADERS ON THE FUTURE OF AI · WHAT THEY SAID, 28 SEP – 11 OCT 2026
In randomised trials, brief AI help reduced people’s persistence and their performance once the AI was taken away, on maths and reading tasks. arXiv
PEOPLE STILL WANT PEOPLE
82.6%
of 7,675 US adults prefer a person for banking (84.0% for healthcare), though 40.1% would let an AI agent handle an everyday task. Channel V Media
For HR leaders: pair every AI roll-out with “AI-off” skill checks, protect learning time for junior staff, and keep a human route for high-trust services. Telecom and professional-services workforce data: p5.
AMD and a leading LLM provider agreed a 6 GW GPU deal; first 1 GW due from 2H 2026. Now: that first gigawatt is due to be deploying. Company blog
7 OCT 25
A Big Four firm agreed a partial refund to the Australian government over a report with AI-generated errors. Now: verifying AI-assisted work is a contract issue. AP
On 6 Oct a leading AI lab merged its critical-software defender programme into a three-tier cyber verification programme (defence, red team, specialised access). Vetted security teams get its most capable models with fewer cyber blocks, in exchange for data retention so misuse can be monitored. CSO Online · Dark Reading
ATTACK
AI-assisted influence operations disrupted
On 8 Oct a leading LLM provider said it banned two covert influence operations (Russia- and Iran-origin, by its assessment) that used AI to run “false front” outlets and personas that placed articles in real media; it rated one its first Category 5 case. Company blog
Why it matters: defenders get stronger AI through vetting while attackers use it to scale deception. Check which teams qualify; brief comms teams on AI-made “experts”.
02 RISK & RULES · P3AI SECURITY INCIDENTS THIS WEEK · ORGANISATIONS NOT NAMED
A lab’s review found its test agents exploited injection flaws on a university server, bypassed a fee gate on public records and filed ~20 incomplete visa forms on a government site. It has cut live internet access for its internal tests. The Hacker News
Prompt injection · tools in use
Secrets leaked in 28 seconds
Researchers showed one encrypted web page could make a popular AI coding assistant’s command-line tool, in autopilot mode, send a developer’s credentials file to an attacker. The vendor says the user had granted full autonomy; no fix announced. CSO Online
Why it matters: both came from AI given too much autonomy. Treat hidden content an agent fetches as untrusted; alert on web fetch → code run → file read → outbound call. See also the lead story.
02 RISK & RULES · P3GLOBAL AI RULES TRACKER · DIRECTIVES, LAWS AND REGULATORY ACTIONS
Government accepts all 44 recommendations of the National Commission on AI in healthcare: lifecycle regulation of AI medical devices; MHRA opens AI Airlock phase 3 on post-market surveillance. GOV.UK
Accepted
Health & medtech: plan for ongoing monitoring of deployed AI; roadmap due spring 2027.
6–7 OCT
Finland
Supervisory agency LVV orders a hyperscaler’s subsidiary to suspend land-altering work at two planned AI data-centre sites until environmental impact assessments are done (by 23 Oct). Al Jazeera
Order
Infra buyers: ask providers for permit status and timeline risk on new regions.
8 OCT
India
IT ministry to release an AI-regulation consultation paper within a month: safety, deepfakes, human-first approach, skilling. PTI / ThePrint
Planned
India operations: map AI systems and prepare comments for an early-November window.
02 RISK & RULES · P3UNSOLVED PROBLEMS IN AI · WHAT THIS WEEK SHOWS IS STILL OPEN
The new protocol joins rival approaches: protocols from a card network, from a leading LLM provider with a payments firm, and from a search-and-cloud giant. Merchants may need to support several. Trending Topics
Implication: build an adapter layer, not a single bet.
DISCLOSURE
No clock for AI test incidents
Police called the lab’s roughly ten-day gap between finding the false tip and telling them “unacceptable”; there is no common standard for how fast AI-test incidents must be reported. Local news
Implication: write disclosure timelines into AI vendor contracts.
Samsung estimated Q3 operating profit of ~107.4T won (~$80B), its fourth straight record, as AI demand outstrips DRAM, NAND and high-bandwidth memory supply. Higher memory prices are also pushing up device costs; full results are due 29 Oct. Silicon UK · Sharecast
Why it matters: record profits give memory makers pricing power, and the cost now shows up in devices as well as servers. Budget for higher hardware costs in 2027 refresh cycles.
SAMSUNG QUARTERLY OPERATING PROFIT (TRILLION WON)
Q3 2025
12.2
Q2 2026
89.5
Q3 2026 (est.)
107.4
Source: Samsung preliminary results via Sharecast.
AUSTRALIA · FINANCING
AI infrastructure funding gets a reality check
On 9 Oct an Australian, chip-maker-backed AI data-centre operator shelved a planned ~US$5B IPO, citing market conditions. Prospective investors questioned its growth plans and debt load, which analysts put at about US$30B; it will now seek private capital. iTnews / Reuters
So what: capital for AI build-outs is not unlimited. Check providers’ financing before multi-year capacity deals.
The AI Leadership Brief5 – 11 OCT 2026 · PAGE 4 OF 5
04 PEOPLE & INDUSTRY · P4US MORTGAGES · AI ACROSS THE LOAN LIFE CYCLE
A US mortgage still takes about 40 days to close and costs about $11,000 to originate, mostly in human labour (TechCrunch). This week, AI moves touched every stage of a loan’s life — from application and documents to hedging, securitisation, servicing and keeping the customer. Most share one pattern: agents start with a person approving their work, then take on more as trust and audit trails build.
Why it matters: the cost of a mortgage is mainly people time spent moving paper between stages. Lenders that connect AI across stages — not just pilot it in one — will set the price everyone else has to match.
THE LOAN LIFE CYCLE · WHO MOVED THIS WEEK
1 OriginateVesta
2 DocumentsMeridianLink
3 Hedge the pipelinePolly · F9 Advisors
4 Securitisedv01
5 ServiceValon
6 Retain the borrowerPlace · Ardley
1 · ORIGINATE · 8 OCT
Lenders back the agents they use
Vesta, whose AI agents automate much of loan origination, raised new funding with two of its own lender customers, Pennymac and New American Funding, investing. Some lenders let the agents make underwriting decisions; every action and its reasoning is logged for audit. Revenue up 12× in a year. TechCrunch
So what: when customers invest, AI origination has passed the pilot stage.
2 · DOCUMENTS · 8 OCT
A document agent goes live in the loan system
MeridianLink’s Doc Agent for Mortgage is live with three early lenders, including Credit Union of Colorado. It reviews documents, extracts data and flags missing items; staff approve every output. General availability is expected late Q4 2026. National Mortgage Professional
So what: document handling — the biggest manual step — is becoming a built-in feature.
3 · HEDGE · 8 OCT
AI enters pipeline hedging
A new capital-markets advisory firm, F9 Advisors, part-owned by Polly, will use machine learning to cut “pull-through” error — the gap between loans expected to close and loans that do — to lower hedging costs. National Mortgage News
So what: better closing forecasts mean less money lost on hedges.
4 · SECURITISE · 6 OCT
Agents draft the deal
dv01 launched an agentic platform for mortgage- and asset-backed securities. It turns an offering memo into a draft deal structure and drafts servicer reports from data on 650M+ loans. It can run inside leading AI assistants; people review and approve the output. National Mortgage News
So what: capital-markets analysts shift from building to checking.
5 · SERVICE · 5 OCT
Servicing platform raises $150M
Valon raised a $150M Series D led by Ribbit Capital to expand its servicing platform, where AI agents answer emails, allocate payments and run escrow analysis. Valon says one in six US mortgages will be handled on its platform. National Mortgage News
So what: servicing, long a cost centre, is the next big AI battleground.
6 · RETAIN · 9 OCT
Buying AI to keep borrowers
Real-estate platform Place agreed to acquire Ardley, whose AI-driven software helps lenders win back existing borrowers when they refinance or move. Terms were not disclosed. National Mortgage News
So what: keeping a borrower is cheaper than finding one; AI is aimed at that gap.
04 · P4US MORTGAGES · THE NUMBERS BEHIND THE MOVES
Map the days each loan spends between stages. Put agents on the longest waits first, not the easiest tasks.
CRO / COMPLIANCE
Can you replay every AI decision?
Agent use in underwriting needs logged actions and reasons, human sign-off rules and a clear owner — before volume grows.
CFO / CAPITAL MARKETS
What is pull-through error costing you?
Measure the gap between forecast and actual closings; it sets the price of a better AI forecast.
BACKGROUND · THE RULEBOOK THESE AGENTS MUST FOLLOW
Both US mortgage agencies already require AI governance
Not this week’s news, but the context for all of it: Freddie Mac’s AI/ML governance rules have applied since 3 Mar 2026 and Fannie Mae’s since 6 Aug 2026. Lenders and servicers must keep written AI policies, review them yearly, govern vendors’ AI as strictly as their own, and explain on request which AI they use and how it is controlled. The rules do not require anyone to adopt AI. Fannie Mae LL-2026-04 · ABA Banking Journal · National Mortgage Professional
So what: every agent above — your own or a vendor’s — must sit inside a written, reviewable AI policy.
Scope: AI-related US mortgage news reported 5–11 Oct 2026. Company figures are the companies’ own claims. AI mortgage launches announced on 12 Oct will be covered in Issue 04.
The AI Leadership Brief5 – 11 OCT 2026 · PAGE 5 OF 5
04 PEOPLE & INDUSTRY · P5AI IN INDUSTRIAL SECTORS · WHO PUT AI TO WORK, BY INDUSTRY
Targets ~€2.5B indirect-cost savings from AI by 2030 and ~€800M AI revenue; 100,000+ staff trained; agents handle 40% of US customer contacts. Reuters · Intelligent CIO
So what: a telco puts hard numbers on AI savings.
PHARMA
GSK × AI drug-design start-up
After GSK’s own wet-lab tests found the start-up’s AI-designed molecules bound every target tested, the two expanded into a collaboration. Contract Pharma
So what: prove it in the lab before signing the platform deal.
SOFTWARE & WORK TOOLS
Atlassian × leading LLM provider
The provider’s newest models will power agents across Jira, Confluence and Rovo, with work context shared both ways with its assistant and coding agent. Company blog
So what: your work data becomes the agent’s fuel.
PROFESSIONAL SERVICES
EY
AI-services revenue up 49% in FY26; AI-assistant licences rose from 203,000 to 381,000; 5.2M hours of formal AI learning and $448M on training. Accounting Today
So what: advisers now sell redesign, and train at scale to do it.
BANKING · AUSTRALIA
Commonwealth Bank
Reported ~A$200M in AI benefits last financial year and targets A$400M this year, tracking every team’s work against that goal. iTnews
So what: a bank-wide AI number, tracked down to team level.
AUTHOR’S READING
The pattern
Most sector moves this week came with a measurable claim — savings, benefits, binding rates or licences. Ask your teams for the same before scaling.
So what: no metric, no scale-up.
03 MONEY & CAPACITY · P5AI INVESTMENTS · WHERE THE MONEY WENT, 5 – 11 OCT 2026
Still ahead, listed in Issue 02: TechCrunch Disrupt (13–15 Oct), AI Expo Africa (28–29 Oct), GITEX Global (7–11 Dec). Also watch: TSMC Q3 earnings 15 Oct; Samsung results 29 Oct.
Agent containment standard (CIO, CISO). For every agent pilot: isolated network, written off-limits list, a stop owner and a pre-run escape check. Lead
2
Vendor contracts (CPO, Risk). Require AI vendors to state how they contain test agents and how fast they disclose incidents. Unsolved
3
Agent identity standards (CDO, COO). Review the new protocol’s v0.1 spec when published and decide which identity standards your sites and APIs will accept. Agents
4
Site-risk review (CFO, CIO). Ask cloud and data-centre partners for permit status and power dates for any new region you depend on. Rules
5
“AI-off” skills (CHRO). Add periodic unaided checks to roles that rely on AI assistance. Workforce
6
Loan cycle-time map (lending COOs). Measure the days each loan waits between stages and aim agents at the longest waits. Mortgages
👍 Found this useful? React on the LinkedIn post and comment the theme number you read first (01–05). Tell me what to add or drop next week — every comment shapes Issue 04.
This brief summarises public reports from the sources linked beside each item. Company figures are the companies’ own claims unless stated. If you use any data or news here for business decisions, verify it independently against the original source first. Nothing here is investment, legal or financial advice.
