November 6, 2026
Sign in to LinkedIn, then select Subscribe.
FOUNDATIONAL UPDATES
OpenAI launches GPT-6 Cyber, a gated model for authorized vulnerability work On November 5, OpenAI introduced GPT-6 Cyber, a cybersecurity-specialized variant of GPT-6 Astra built for authorized vulnerability discovery, exploit development and web penetration testing. On ExploitGym it scored 48.5%, versus 32.1% for GPT-5.6 Cyber, about a 51% relative improvement. OpenAI classified it Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. Access runs through Daybreak Red with identity checks and organization security controls, and starts with select approved Daybreak Red customers. Researchers used it to develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity Firefox vulnerabilities. Mozilla released fixes for many of those. Early customers and quotes include Cisco, Palo Alto Networks, CrowdStrike, CodeRabbit and others. On alignment tests, it stayed in scope 99.93% of the time without a deployment prompt and 100% with one on ExploitGym Honeypot.
Why it matters for pharma: Regulated clinical, manufacturing and patient systems need stronger defensive AI security tools, and they need governed authorized use. Access is gated. Ask IT/security and Quality how they'd use Daybreak-class tools inside GxP environments without going out of scope.
Source: OpenAI, "Introducing GPT-6 Cyber" https://openai.com/index/introducing-gpt-6-cyber
Google makes Gemini Omni 1.2 Flash generally available, with longer video and a price cut On November 5, Google said Gemini Omni 1.2 Flash is generally available on the Gemini API and on Gemini for business / Google Cloud. Native video length goes up to 15 seconds, from 10 with Omni 1.1. It supports multi-turn conversational edits. Price is down 30% versus Omni 1.1, to $0.07 per second for 720p. Outputs carry SynthID watermarking, and Provisioned Throughput is available on Google Cloud.
Why it matters for pharma: Cheaper enterprise video helps medical education, patient materials, site training and commercial content. Watermarking and Cloud GA matter for brand safety and procurement. You still need human review of claims in regulated materials.
Source: Google, "Introducing Gemini Omni 1.2 Flash" https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash/
CLINICAL TRIALS & AI
PhaseV launches BonV, an agentic layer that turns clinical questions into decision memos On November 3, PhaseV launched BonV, an agentic layer on its causal-AI platform. Teams ask plain-language questions. BonV routes them through modules for causal modeling, target trial emulation, ClinOps Optimizer and Trial Optimizer, then returns a documented decision memo. PhaseV says in one trial four modules cut costs $32M and shortened duration by eight months. In another, the company says BonV supported due diligence for a multibillion-dollar metabolic/obesity deal in two weeks. Both figures are company-stated. BonV is already with select pharma sponsors, with a phased enterprise rollout supported by Forward Deployed Engineers. PhaseV says it has supported 100+ programs for 50+ sponsors, including 10 of the top 20 pharma companies.
Why it matters: Turning a protocol or portfolio question into a documented memo is the kind of work ClinOps and clinical development already do by hand. But treat the company-stated savings as a case study to pressure-test, not a guarantee. And keep a human as the decision owner. The memo is only as good as the question and the data behind it.
Source: PR Newswire / PhaseV https://www.prnewswire.com/news-releases/phasev-launches-bonv-an-agentic-ai-orchestrator-turning-clinical-questions-into-defensible-decisions-302927742.html
nPhase raises $30M Series D to expand its AI-native clinical data platform Florida-based nPhase closed a $30M Series D led by Viking Global Investors, announced November 2 (Fierce Biotech covered it November 4). It's the company's first equity raise since 2017. The funds expand nPhase.ai, its Unified Data Management System and Statistical Computing Environment, from study startup through regulatory submission. The platform uses agentic AI for discrepancy flagging and draft statistical programming, with humans keeping decision authority. nPhase says its SCE aims to compress submission-prep workflows that often take four or more months into weeks. The company also operates REDCap Cloud EDC.
Why it matters: Submission prep is where a lot of programs bleed calendar time. An SCE that drafts statistical programming under human authority could help, if the audit trail and validation story hold up. So sponsors comparing vendors should ask how the agentic steps are logged, who signs off, and whether the system has been used in a real submission package.
DRUG DISCOVERY & AI
Addgene connects its plasmid and viral prep catalog to Claude for Life Sciences On November 5, Addgene announced a read-only MCP connector for Anthropic's Claude for Life Sciences. Researchers can search more than 177,000 plasmids and about 1,000 viral preps, pull item details and get GenBank download links. The connector cannot place orders or change accounts.
Why it matters: Lab and discovery teams already spend time hunting reagents. A search that sits inside the chat tool they use for protocol design is a small, practical win. Because it's read-only, the governance bar is lower than for anything that can spend money or change an account.
Source: GlobeNewswire / Addgene https://www.globenewswire.com/news-release/2026/11/05/3396289/0/en/addgene-integrates-plasmid-and-viral-prep-catalog-into-anthropic-s-claude-for-life-sciences.html
Anthropic team uses Claude to solve the first 3D structure of a prokaryotic vault particle On November 4, Anthropic published research on the first 3D structure of a prokaryotic/phage vault particle. The L05 phage protein forms a particle of 96 copies, about 9.5 megadaltons and roughly 54 by 45 nm. Isolation took about 90 minutes versus a roughly three-day conventional prep. The first full structural model took about 34 hours. The team estimates that would normally take weeks. They also engineered larger vaults with duplicated wall segments, with about 20% more enclosed volume. A preprint was submitted to bioRxiv. The function of vaults is still unknown. One possible biotech use is as enzyme-delivery nanocapsules. Claude helped search databases, design constructs, optimize isolation and process cryo-EM data. The team also notes Claude mistakes, including a wrong growth medium additive that cost two weeks and a skipped cryo-EM step.
Why it matters: This is a concrete example of a frontier model speeding wet-lab structure work, with the mistakes left in the write-up. Discovery and CMC teams thinking about AI in the lab should copy that honesty. Track what the model got wrong, not just the clock time it saved.
Source: Anthropic https://www.anthropic.com/research/investigating-vault-particles-with-claude
DEALS & PIPELINE
OpenAI-backed Biossil buys Theriva's antibiotic-degrader portfolio for up to $91M in milestones On November 2, Theriva Biologics sold SYN-004 (ribaxamase), SYN-007 and certain SYN-006 assets to Biossil Labs, an AI-native company with Toronto and Boston ties and backing from OpenAI and Founders Fund (per Fierce Biotech). Terms include $350k upfront, up to $3M in regulatory milestones, up to $88M in sales milestones and low-single-digit royalties capped at $250M. Theriva frames total potential as up to $91M in regulatory and sales milestones. Biossil assumes all future development costs, including the paused Phase 1b/2a HCT/aGVHD trial NCT04692181.
Why it matters: An AI-native buyer picking up a paused anti-infective asset is a signal that capital is looking for programs where modeling and trial design can restart cheaply. For BD teams, it's another data point on how AI-native shops price early clinical risk.
Source: Theriva Biologics https://therivabio.com/press_releases/theriva-biologics-announces-sale-of-syn-004-ribaxamase-and-related-antibiotic-degrading-enzyme-portfolio-to-biossil-labs-inc/
REGULATORY & AI
Study finds AI cited in just 7 FDA warning letters since 2022 A study by Kwiecinski and Nandy, published October 31 and covered by FDAWeb on November 4, analyzed 3,011 FDA warning letters from January 1, 2022 through September 28, 2026. Seven letters (0.23%) referred to AI: three device, three tobacco and one drug. Only one, an April 2026 letter to Purolea Cosmetics Lab, treated AI use itself as a violation. That case involved AI-generated CGMP documents without quality-unit review under 21 CFR 211.22(c). The authors say FDA mostly applies existing rules, not AI-specific enforcement.
Why it matters: The lesson is familiar. FDA cares that your Quality Unit reviews the documents, not that you used AI to draft them. If AI writes SOPs, batch records or other CGMP docs, the quality-unit review has to be real and documented.
Source: FDAWeb https://www.fdaweb.com/ai-cited-in-just-7-warning-letters-since-2022-study/
FDA webinar on final ICH M15 stresses flexibility on model evaluation FDA held a stakeholder engagement webinar on October 29 on the final ICH M15 guideline (General Principles for Model-Informed Drug Development), which FDA implemented in June 2026. FDAWeb summarized takeaways on November 3. Model analysis plans are good practice but not required in advance. High model risk or impact is about evaluation rigor and early engagement, not automatic rejection. Failing predefined validation criteria does not automatically disqualify a model if it is still fit for the question of interest and context of use. FDA reviewer Xiaofei Wang urged early MIDD communication documents when modeling aims to reduce animal studies.
Why it matters: Sponsors using models to cut animal work or shape trial design should talk to FDA early, and write the context of use clearly. Because a failed validation criterion is not an automatic dead end if the model still answers the right question. That's useful flexibility, and it puts more weight on how you frame the question up front.
Source: FDAWeb https://www.fdaweb.com/ich-m15-model-evaluation-flexibility-submission-requirements/
Originally published on LinkedIn. Read the original
