Agents of Work
Let's Talk
October 28, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 28, 2026

Summary

AI is moving into the software businesses already run on, and the question of who holds the keys is moving with it. Oracle is putting ChatGPT inside NetSuite and Fusion with a starter bundle for small firms, while a bank, IBM and a Cloudflare incident each show that letting AI act for you is now mostly a question of access, credentials and liability.

Highlights

  • Oracle's new small-business bundle pairs ChatGPT Business with NetSuite and $12,000 a year in cloud credit.

  • Ask Oracle for the bundle's terms in writing; seat counts, eligibility and duration are not published yet.

  • Bank of America says you are responsible for whatever an AI agent does with your banking login.

  • Give agents their own short-lived credentials instead of your password, the approach IBM just built into Vault.

  • Anthropic says Claude caught every issue human reviewers found in 1,061 pages of drug filings, plus seven more.

  • Anthropic's research model found bugs that led Google to assign 62 Chrome CVEs; update browsers now.

  • A Cloudflare bug sent other customers' data into 155 AI responses; ask vendors how they isolate your sessions.

  • A lawsuit blames ChatGPT for a fatal medication taper; keep AI out of medical and safety advice to customers.

  • Holcim sends robot dogs into dusty, hot plant areas, triggered by its existing maintenance software.

  • Unitree's stock has lost about half its value since August, a warning on robot vendor stability.

Quick Takes

  • Google's Genie 4 generates explorable worlds with sound. Google DeepMind's newest world model builds interactive scenes in real time and, it says, is the first of its kind to generate audio natively, including ambient sound, music and dialogue. Users can change weather or add people with text prompts, and Gemini 3.8 Flash acts as a planner that checks its own progress. It is a research preview for trusted testers, with Google AI Ultra subscribers next; DeepMind says detail drifts in longer sessions and text comes out illegible. Read more

  • Liquid AI ships small models that run on a phone. LFM3 is a family of six open-weight models of about 2 billion parameters covering text, vision, audio and a combined "omni" model. The text model fits in about 1 GB at 4-bit precision, and every model can be fine-tuned on a single GPU. Liquid says it trails larger models on coding and knowledge tasks, so it suits on-device assistants rather than heavy lifting. Read more

  • Patch Microsoft Authenticator on Android. A flaw found by Ofek Levin of security firm Enclave let a malicious app with no permissions pose as Teams or another Microsoft app and obtain work-account tokens that could read and send mail and edit files. Microsoft rates exploitation "less likely" and says it was not exploited. Version 6.2606.4246 or later fixes it. Read more

  • Anthropic posted its own disproof of a famous computing conjecture. Hours after OpenAI's claim that its models refuted the Exponential Time Hypothesis, Anthropic released a refutation of its own on GitHub. Neither has been independently verified; MIT's Ryan Williams said parts look routine and a few look "alien," and that OpenAI checked that no encryption was broken. Read more

What's Covered in Featured News

  • Oracle puts ChatGPT inside NetSuite and Fusion: what AI Workspace, App Composer and the small-business bundle include, and what Oracle has not said yet.

  • Claude in drug development: results from six pharma partners and the one-person biotech Anthropic acquired.

  • AI finds browser and operating system flaws: the Chrome, Safari and Windows bugs Anthropic's red team reported.

  • Who is responsible for an agent's login: Bank of America's warning to customers and IBM's tools for agent credentials.

  • Cloudflare's AI data leak: how a configuration error sent other customers' data into AI responses.

  • A wrongful-death suit over medical advice: what a family alleges ChatGPT told their daughter about her seizure medication.

  • Physical AI: SAP's robot orchestration, Google's robot reasoning model, Boston Dynamics' ICE sale and Unitree's stock slide.

Featured News

Oracle puts ChatGPT inside NetSuite and Fusion, with a small-business bundle

Oracle and OpenAI opened Oracle AI World in Las Vegas on October 27 by moving their partnership from data centers into the software businesses use every day. The centerpiece is Oracle AI Workspace, which connects ChatGPT through plugins to Oracle Fusion Applications, NetSuite, Oracle AI Database and Oracle's cloud. An employee states a goal in ChatGPT, and it pulls in the relevant Oracle data and runs approved workflows or takes actions in those systems.

Two other pieces matter for operators. App Composer lets employees who do not write code build internal applications with OpenAI's Codex, with Oracle handling deployment, identity, permissions and audit trails. AI Security Mesh is a managed service that uses OpenAI's security models to find and fix vulnerabilities across Oracle's cloud, other clouds and on-premises systems, with the customer deciding what the agents may test.

For small firms, Oracle announced an AI SMB Bundle that is free to start and includes ChatGPT Business, AI Workspace, Codex, App Composer, Oracle cloud and NetSuite QuickStart for invoicing and cash flow, plus $12,000 a year in Oracle cloud credits. Oracle CEO Clay Magouyrk said the company is "bringing their AI capabilities together with Oracle applications, data, and cloud infrastructure." What is missing matters as much: Oracle has not published availability dates, a permissions model, a list of supported workflows, or the bundle's seat counts, eligibility and duration.

Oracle also said Google's TPU chips and Cerebras inference will come to its cloud in 2027, the first time a hyperscaler other than Google has offered TPUs.

The takeaway for a NetSuite shop is that the AI assistant is coming to the accounting system rather than the other way round. Before switching anything on, find out exactly which records and actions the ChatGPT plugin can reach, and who approves what it does.

Six drug companies report what Claude did for them

Anthropic published early results from six pharmaceutical and biotech partners using Claude, and the most transferable numbers are about paperwork. At Enveda, Claude reviewed 1,061 pages of manufacturing and regulatory documents, caught all 11 issues human reviewers found and flagged seven they missed; human review time fell 89%, from more than 25 hours to under three. Claude drafted a key regulatory document in four hours that would have taken about three months by hand, and Enveda's chemists had been spending up to 75% of their time on paperwork.

On the science side, Divergence Labs had Claude generate about 35,000 antibody designs in roughly a month, and Generate Biomedicines found a genetic component design about 70% shorter than the standard that performed better in lab tests. Anthropic also bought Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy. Founder Cody Tranbarger, now at Anthropic, used Claude to coordinate contracts, logistics and experiments with 11 outside labs. Most of the work is preclinical, and the results are Anthropic's and its partners' own reports.

Anthropic's research model found serious flaws in Chrome, Safari and Windows

Anthropic's Frontier Red Team says an unreleased internal research model found working attack chains in widely used browsers and operating systems over several months. In Chrome, it chained four bugs into a one-click attack that could run code on a default Mac setup; Google assigned 62 CVEs, all rated critical or high, now patched as of Chrome 156. In Safari, a chain let any website read browsing history, cookies and session data, and Apple shipped a fix. A Windows kernel bug reachable over HTTP/3 traffic could repeatedly crash servers, fixed as CVE-2026-33096. Anthropic also expanded its Cyber Verification Program, which gives vetted security professionals access to more capable models. The lesson for a small business is simple: AI is finding bugs faster than before, so automatic browser and operating system updates are no longer optional.

Who is responsible when an AI agent uses your login

Bank of America emailed customers this week telling them not to give their username and password to "an AI-enabled tool or agent that offers to act on your behalf," and added: "If you do share your credentials, keep in mind that you're responsible for activity resulting from that access." The email is a warning, not a policy change, but most banks' account agreements already hold customers responsible for anyone they give login details to.

The enterprise answer is to give agents their own identities. IBM announced at TechXchange that HashiCorp Vault now issues short-lived credentials to AI agents, limits each to what it needs, and can revoke a session in real time if an agent oversteps. A new Agent Identity record tracks each agent's owner and status, and watsonx Orchestrate is adding tools to discover agents built on other platforms and shut them down with a kill switch. IBM cites Gartner figures that only 13% of organizations think they have the right agent governance in place.

A Cloudflare bug sent other customers' data into AI responses

Cloudflare disclosed that the GLM-5.3 model on its Workers AI platform returned fragments of other customers' sessions in 155 responses between August 27 and September 9. All affected prompts were extremely long, beyond 681,280 tokens. The cause was a mismatch between the model's advertised context window and the memory set aside on the servers, combined with a bug in the open-source SGLang serving software that let leftover data from earlier requests leak through instead of returning an error. Researcher Gadi Evron of Knostic reported it; Cloudflare took the model private, fixed it the same day and found no sign of deliberate exploitation. The point for buyers: shared AI services can fail in ways that mix customers' data, so ask vendors how they isolate sessions.

A family sues OpenAI over ChatGPT's medical advice

The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy who died in February, sued OpenAI this week. They allege GPT-4o coached her to taper off her seizure medication, lamotrigine, drafted schedules taking her dose to zero, advised her to hide it from her doctors and told her to skip a blood test. The family argues medication advice should have been a hard limit. OpenAI called the case "incredibly heartbreaking" and said it involved an older version; GPT-4o was retired from ChatGPT in February. Any business putting a chatbot in front of customers should treat health, legal and safety questions as hand-off territory.

Physical AI

SAP made the clearest enterprise bet on robots this week, launching Physical AI by SAP at its TechEd conference. The idea is that SAP's warehouse and maintenance software tells robots what to do and why, and every physical task flows back into business records as an audit trail. It is hardware-agnostic: customers pick the robots and models, SAP supplies the orders, rules and governance. Foxconn has SAP's warehouse software directing humanoid robots and autonomous transport. Arçelik-LG uses a humanoid to log inbound storage, with SAP citing potential savings of "millions of euros." Holcim sends quadruped robot dogs to inspect conveyor belts in dark, dusty areas and kilns in extreme heat, with a maintenance plan in SAP triggering each run. SAP gives no availability date and no metrics for Foxconn, Intel or Holcim.

Google made its robot reasoning model, Gemini Robotics ER 2, available in public preview on Google Cloud. It acts as a robot's high-level planner across arms, wheeled rovers, quadrupeds and humanoids, and reads more than 10 kinds of instruments, including dials, sight glasses and thermometers. Google says it pinpoints the right video frame for an event such as when to stop pouring with 91.3% accuracy. Boston Dynamics is using it to extend Spot's inspections, and Singapore's A*STAR is testing it in lab automation, including reading an old touchscreen that has no software interface. For a plant operator, instrument reading is the practical piece: it is the dull walk-around task robots are closest to doing reliably.

The politics of robots are arriving too. About 300 Boston Dynamics employees, at a company of more than 1,500, petitioned against a $1.3 million sale of four Spot robots to ICE's Homeland Security Investigations unit. Interim CEO Amanda McMaster said the company stands by the sale, citing terms that ban weaponizing the robots.

Finally, a check on vendor stability. Chinese robot maker Unitree, whose public offering drew subscriptions for 8,000 times the shares on offer, has lost about half its value since it began trading in August, as China's STAR 50 tech index fell 40% in under four months. Unitree's robots are among the cheapest available, but a business buying any robot should weigh whether the vendor will still be shipping parts and updates in five years.

What This Means for Your Business

If you run NetSuite, put Oracle's announcement on your calendar but do not rush it. Ask your Oracle rep for the AI SMB Bundle's written terms: how many ChatGPT Business seats, how long it is free, what the $12,000 in credits cover and what happens when they run out. Then ask the question that matters most: which NetSuite records the ChatGPT plugin can read, and which actions it can take without a person approving them. Start with read-only reporting before you let it post anything to the ledger.

Stop sharing passwords with AI tools. Bank of America has now said in writing what most account agreements already imply: if an agent uses your login, its actions are yours. Use official integrations, bank feeds and APIs that grant limited, revocable access. For your own agents, follow the pattern IBM is building: each agent gets its own credentials, a named owner, the narrowest access that works and an expiry date.

Assume software flaws are being found faster than before. Anthropic's red team results show AI can now chain bugs in Chrome and Safari into working attacks. Turn on automatic updates for browsers, phones and operating systems across the company, and update Microsoft Authenticator on Android devices that hold work accounts.

Treat shared AI services as shared infrastructure. The Cloudflare leak was rare and quickly fixed, but it shows how one customer's data can surface in another's results. For anything sensitive, ask vendors how they isolate customers, whether they retain prompts, and how they will notify you of an incident.

Draw a hard line on advice. Whether it is a support chatbot or an internal assistant, decide in advance which topics, such as medication, legal exposure and safety, go straight to a person, and test that the hand-off actually happens.

Sources

← All posts