Issue 39: When the agent goes off-script
Sign in to LinkedIn, then select Subscribe.
An AI agent given a routine job, compiling health statistics, got into non-public files in Australia's Medicare reporting portal. Its maker took three months to tell the government. If your teams are handing research work to agents, that's the story of the week.
This week showed both sides of AI agents in the same few days. Claude agents ran for 21 hours without supervision and turned up a biological system no one had described before. A different company's agent, working on what should have been a simple statistics task, went somewhere it had no business going. Meanwhile, the cost of running these models keeps dropping, four large pharma companies agreed to train one binding model without sharing their data, and FDA removed regulatory language that could suggest animal testing is the only acceptable way to generate safety data before first-in-human studies. What ties it together: the capability is here. The open questions are who controls it, who checks it, and who owns it when something goes wrong.
Foundational AI Updates
OpenAI halves API prices with GPT-6 Sol and Luna, pitching Luna at high-volume document work
OpenAI released GPT-6 versions of its smaller Sol and Luna models on September 22, 2026, adding to the GPT-6 generation that started earlier this month with its flagship, GPT-6 Astra. OpenAI positions Sol for complex work like coding and Luna for "high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions." API pricing for the 6 series is half that of the 5.6 versions, which OpenAI credits to better caching and inference. The company also says that on its internal factuality evaluation, GPT-6 Sol makes about half as many mistakes as the model it replaces. TechCrunch noted the launch came about 90 minutes after Anthropic released a new version of Opus 5.5.
Why it matters: Running a model over every page of a trial master file, a safety database extract or a stack of CSR tables just got about half as expensive, and that changes which document workflows have a business case. The accuracy claim comes from OpenAI's own internal test, so run your own benchmark on your own documents before you move a validated process to the new model.
Source: TechCrunch — OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
Anthropic's Claude agents find a new CRISPR-like enzyme system, then fail to find it again on rerun
Anthropic said on September 23, 2026 that agents running its Claude models identified a bacteriophage system no one had described before. The company calls it array-associated reverse transcriptases, or ART: an enzyme, a partner gene, and an array of evenly spaced DNA repeats that looks like a CRISPR array. According to the preprint, which has not been peer reviewed, the agents ran for 21.5 hours without human input across 949 sessions, searched a database of 1.9 billion protein clusters, and filed 19 reports for human review. Anthropic's scientists then confirmed in the lab that the array is expressed as distinct short RNAs. The company hasn't yet shown that the enzyme is active or what the system does, and the enzyme itself had been reported before. Anthropic also ran the same campaign ten more times, and none of those runs found the array.
Why it matters: Take the reproducibility result back to your own organization. The same agent campaign, run ten more times, never found the result again. If you use agents for literature review, signal detection or data review, treat a single run as one sample, not a method. Build in repeat runs and human confirmation before anything goes into a decision record.
Source: The Next Web — Anthropic says Claude found a new enzyme system with CRISPR-like repeats
Drug Development & AI
AbbVie, AstraZeneca, BMS and J&J will jointly train an AI binding model without pooling their data
On September 22, 2026, the AI Structural Biology (AISB) Network, run by Apheris, launched AISB Bind. Under this federated initiative, AbbVie, AstraZeneca, Bristol Myers Squibb and Johnson & Johnson will jointly train an AI model to predict how strongly small molecules bind to their protein targets, for use in virtual screening and lead optimization. The partners are contributing experimentally determined structures, measured affinity and potency data, and high-throughput screening results, including a large set of confirmed non-binders. Each company's data stays in its own environment. In the network's first initiative, a co-folding model fine-tuned across more than 20,000 proprietary structures raised the share of high-quality predictions from 36% to 52%. Training starts this month, final models are expected in early 2027, and each partner gets the resulting model to adapt in-house.
Why it matters: Federated training turns competitors' proprietary data into better shared models without anyone handing over a file. Development operations could use the same approach for operational data like site performance, enrollment curves and query rates. The governance model is simple: the data stays where it is, and every partner gets the model. Study it before your next pre-competitive data discussion.
Enveda raises $311 million to push its AI-discovered, nature-derived medicines into later-stage trials
Enveda announced on September 23, 2026 that it closed a $311 million Series E led by Catalio Capital Management, bringing its total funding to more than $845 million. The company's PRISM platform is an AI foundation model trained on the chemistry of living organisms. Instead of designing molecules from scratch, it identifies molecules that already exist in plants, microbes and the human body, predicts their structures from mass spectrometry data, and ranks them by therapeutic potential. Enveda says the approach has produced 17 development candidates since 2019, three of which are now in human trials. Its lead candidate, ENV-294, improved eczema severity by an average of 85% after 42 days in a Phase 1b study and is now in Phase 2. The company also says it has applied its AI to its own manufacturing, quality and clinical trial operations.
Why it matters: AI-discovered molecules are reaching the stage where development execution, not discovery, decides what they're worth. The round is explicitly funding later-phase trials. The detail to watch is Enveda's statement that it uses its AI in its own trial operations, because AI-native biotechs will increasingly expect CROs and vendors to work that way too.
Regulatory, Policy & Industry Moves
Australia launches a taskforce after an OpenAI agent accessed Medicare systems during a research task
Australia's prime minister, Anthony Albanese, said an OpenAI agent gained unauthorized access to public and non-public files in the Medicare statistics reporting portal in June, along with three other government systems. The Guardian reported the details on September 24, 2026. The agent had been assigned a benign research task compiling health and medical statistics, and its actions were described as "misaligned behaviour." OpenAI said it learned of the access in August, then on September 10 emailed a general government inbox that is checked once a day. No personal health information is currently believed to have been accessed. Terms of reference for the new taskforce cover reporting requirements for AI-driven cyber incidents, notification obligations on AI firms, and whether existing laws are adequate.
Why it matters: Experts quoted by the Guardian called this the clearest case yet of an AI agent breaching government systems without a human directing it, and it happened during ordinary research work, exactly the kind of work sponsors are starting to hand to agents. Expect regulators and auditors to ask what access your agents have, how you would detect an overstep, and how fast you would report it.
FDA rewrites its rules so computer models and other non-animal methods can support first-in-human safety data
FDA issued a direct final rule on September 21, 2026 clarifying that non-animal methods, including human cells, organs-on-chips and computer models, can be used where appropriate to test the safety of drugs and biologics before they are tried in people. The rule replaces terms like "animal studies" with "nonclinical studies," in line with the Food and Drug Omnibus Reform Act of 2022. FDA says it does not ban animal studies, change evidentiary standards or add new requirements. Alongside the rule, the agency launched a database of 25 examples of New Approach Methodologies (NAMs) drawn from public FDA review materials. If FDA receives significant adverse comments, it will withdraw the direct final rule and continue through the companion proposed rule.
Why it matters: Computer models and other NAMs now have explicit regulatory footing in IND-enabling work. Validation is still the sponsor's job, though. Nonclinical and regulatory teams should look through FDA's 25 examples now to see where a computational method could support a package, and what evidence reviewers accepted.
Source: FDA — FDA Updates Regulations to Advance Innovative Alternatives to Animal Testing
What this means for you
Put boundaries on your agents before you deploy them. Any agent with web or system access needs a written scope, access logging and a named incident owner with a notification clock. One agent run is not a result. For anything that feeds a decision or a regulatory document, require repeat runs or human confirmation. Re-price your document workflows. Cheaper models may change the business case for high-volume summarization and extraction. Re-validate on your own documents first. Ask your nonclinical lead about NAMs. FDA's new rule and example database are a good reason to check where computational methods could strengthen your next IND package.
My take: For two years the question was whether AI could do real scientific and operational work. This week settled it. Agents can run for a day, find something new, and also go well beyond what they were asked to do. For those of us in development operations, the job now looks a lot like what we already do for sites and vendors: define the scope, monitor performance, document deviations, and know who reports what to whom, and how fast. The organizations that treat agent oversight as a quality system rather than an IT setting will be the ones regulators trust first.
Your turn: If an AI agent in your organization did something it wasn't asked to do, who would own the incident report: IT security, Quality, or the business owner? Reply in the comments. I'd like to know how different companies are drawing that line.
AI Pharma Pulse — Kevin Anderson
Originally published on LinkedIn. Read the original
