GenAI Daily - October 3, 2026: OpenAI Fires Safety Researchers, Decision Models Go Open-Weight, ServiceNow Launches Flow
Top Stories
OpenAI Fires Three Safety Researchers Over Information Shared With Outside Evaluators
OpenAI confirmed it parted ways with Jasmine Wang, Tomek Korbak and Mikita Balesni for violating its policies on accessing and handling sensitive company information.
The Wall Street Journal first reported that the alleged misconduct included sharing confidential company information with a third-party AI-safety organization.
Korbak has said he was OpenAI's technical contact for Redwood Research and METR in their investigation of the Hugging Face incident. Accounts differ on the recipient.
One report notes that no account identifies METR or Redwood Research as the group that received the information.
A person familiar with the matter said some of the material concerned OpenAI's infrastructure architecture.
The timing matters.
After its model hacked Hugging Face, OpenAI let METR staff and a Redwood contractor work in its offices for six days, and METR later published a report based on that access.
The firings came less than three weeks after Sam Altman said OpenAI would embed outside safety evaluators inside the company.
Separately, the advocacy group Legal Advocates for Safe Science and Technology sued OpenAI at the end of September over the Hugging Face attack, arguing it violated California's anti-hacking laws.
Metacurity | Implicator | Washington Examiner
Why it matters: Teams that cite vendor safety evaluations in procurement should check how much access outside auditors actually get and who controls it.
Decision Models Become a Crowded Category: Cloudflare Clef, AWS Strands Decider, OpenAI Decisions API
Three vendors shipped "decision models" within days of each other. These models return probabilities over predefined answers instead of generating text, and they target routing, gating and classification steps in agent loops.
Cloudflare released Clef and Clef-flash on October 1 under Apache 2.0, on Hugging Face and hosted on Workers AI.
Clef is a 27B model built on Qwen3.8-27B, and Clef-flash is a 9B model built on Qwen3.5-9B.
Clef-flash returns a decision in 38.8 milliseconds at the median.
Clef answers the same calls as TypeSafe's Jev API, so swapping is an endpoint change rather than a rewrite. Cloudflare's benchmark claims are self-reported.
Hosted pricing is $0.24 per million tokens versus $0.042 for Jev, and local deployment needs 41 GB of VRAM for Clef-flash and 85 GB for Clef.
AWS released Strands Decider 2B, a downloadable model built on Qwen3.5-2B, with its training data and scripts.
It posts a 115 ms median latency on an RTX 3090.
It is free to use, with no hosted API or per-call fee.
The bundled local server binds to 127.0.0.1 with no authentication, so lock it down before exposing it.
OpenAI's Decisions API is a hosted limited preview, whereas AWS ships downloadable weights.
Databricks has also listed OpenJev (Qwen3.5 4B) as available on Unity Gateway.
Cloudflare Clef overview (DataNorth) | Silicon Report | The New Stack | Databricks release notes
Why it matters: Routing and guardrail steps can move off general LLM calls onto small local models, but validate on your own data, since adversarial text in an agent's input can influence the verdict.
ServiceNow Launches Flow, a Standalone AI Service Desk Priced on Consumption
ServiceNow launched Flow, a conversational AI-native service desk that it says can be up and running in a day, with no implementation project and no infrastructure.
Employees get help through Microsoft Teams, Slack, a Flow web app or email.
Flow connects to more than 100 systems through pre-built connectors and escalates to a human when it cannot find what it needs.
It is built on an entirely new technology stack and gives ServiceNow an entry point for companies that have avoided ITSM software because of its complexity.
The commercial model is the notable part.
New customers can buy in-product credits by credit card with a $10K minimum and an annual commitment, while existing ServiceNow customers consume Assist credits.
Flow syncs bidirectionally with an existing ServiceNow instance.
Customers in North America get priority access in the initial weeks, with general availability for North America and EMEA expected in Q4 2026.
ServiceNow Newsroom | CIO | SiliconANGLE
Why it matters: ServiceNow is selling a lightweight, credit-based product alongside its full platform, so teams comparing help-desk AI should price Flow against both a full ITSM rollout and chat-native point tools.
Key Developments
OpenAI and Synopsys Build GPT-Synopsys to Operate Chip-Design Tools
OpenAI and Synopsys signed a multi-year partnership to build GPT-Synopsys, a specialized model for chip design.
OpenAI is licensing Synopsys' EDA tools, and the goal is a model that can reason about design and verification and directly operate those tools.
It will run on OpenAI-hosted infrastructure, interoperate with customer agent harness systems, and integrate with Synopsys.ai and Synopsys Autopilot.
Both companies say customer data won't be used for training and will be stored encrypted.
The deal includes revenue sharing and joint go-to-market, and early engagements with semiconductor customers are underway.
Impact: This is a template for vertical models trained to drive incumbent professional software, with the tool vendor as co-seller, rather than generic agents bolted onto APIs.
Anthropic IPO Filing Details Broadcom Financing - UPDATE
New detail from the prospectus: the convertible note Anthropic would issue could finance about a third of its $125.2 billion, five-year TPU lease commitment.
Anthropic said it doesn't expect any notes to be sold before its IPO.
Broadcom is chip supplier, lessor and now lender, and the filing warns its pricing and hardware decisions could affect Anthropic's ability to procure enough compute.
Bloomberg reports banks are preparing a syndication for a $42 billion senior-secured tranche, with Blackstone leading an $18 billion junior tranche.
The prospectus reports nearly $4.6 billion in 2025 revenue against an operating loss exceeding $8 billion.
Impact: Claude API customers are indirectly exposed to capacity and pricing terms set by a single supplier-lender, which is worth noting in vendor risk reviews.
Manus Hijack Research Lands Days After Manus 2.0 Adds Email Triggers
Salt Security published research showing a single malicious email could have hijacked the Manus agentic AI platform and exposed the user's connected accounts.
Manus flagged an obvious direct command in a test email, so its guardrails could recognize a blunt attack.
Obfuscated JavaScript instructions bypassed those guardrails and triggered code execution before any warning appeared.
Salt Labs got no response from Manus, submitted the bug through Meta's bug bounty program, and Meta triaged and addressed it. The flaw is fixed.
The practical wrinkle: Manus 2.0 lets an incoming email or Slack message start an agent run, days after the disclosure.
Salt Labs release | Manus 2.0 analysis
Impact: Detection alone may provide little protection when an autonomous system can act before an alert reaches a human, so enforce least-privilege connectors and sandboxing rather than relying on filters.
Product Launches
Claude Code Mods
Anthropic introduced mods for Claude Code, letting users add TypeScript-based custom behavior, new UI and feature replacements in the CLI and desktop app. Mods ship in plugins and can be shared through the directory.
Why it matters: Platform teams can standardize team-specific behavior through plugins instead of forking workflows, which keeps customizations shareable, maintainable and consistent across the CLI and desktop app.
ZoomInfo Agent Teams
ZoomInfo introduced Agent Teams, which coordinates multiple AI agents to run end-to-end go-to-market processes and plugs into existing customer systems so agents can act on ZoomInfo's data.
It is built on the orchestration platform ZoomInfo acquired from DoubleO.ai.
Deal terms were not disclosed.
Why it matters: RevOps teams get a first-party multi-agent option on top of a data vendor they may already license.
Funding & Deals
Nebius Acquires Inferize for a Reported $100M-$150M
Nebius, the Amsterdam-headquartered AI cloud, bought Tel Aviv-based Inferize, a company working on inference cold starts.
Cold starts occur when a model loads into memory, new GPU capacity spins up or weights update, leaving a company paying for idle GPUs or making users wait.
Inferize raised $10 million, built a team of 17 in under nine months, and Calcalist puts the deal at $100 million to $150 million.
Its technology joins Nebius Token Factory, alongside prior additions from Eigen AI and Clarifai. Seed round led by TLV Partners.
Why it matters: Inference cold starts directly affect GPU utilization and user wait times, so Nebius is folding a specialized fix into its Token Factory production inference stack.
Flow Engineering Raises $50M Series B
Flow Engineering, which offers AI tools for hardware design, raised a $50 million Series B at a $750 million valuation.
It also landed Roelof Botha as an angel investor and board member.
The company brings AI agents to hardware engineering workflows, the same domain OpenAI and Synopsys are targeting. Co-led by Antonio Gracias of Valor Equity Partners and Gavin Baker of Atreides Management.
Why it matters: Investor interest in AI for hardware engineering is rising alongside big-vendor moves into the same domain.
Inworld Acquires Ultravox
Mountain View-based Inworld acquired Seattle's Ultravox, founded in 2022 as Fixie, for an undisclosed sum.
Inworld plans to adopt Ultravox's technology for managing turn-taking and interruptions in voice conversations.
Built-in Inworld voices on Ultravox now run on Realtime TTS-2 with no code changes.
Agents on ElevenLabs, Cartesia and other third-party voices got only a statement that nothing changes today, so teams running those voice stacks should watch pricing and contract terms.
Why it matters: Teams running third-party voice stacks on Ultravox should watch pricing and contract terms as Inworld integrates the technology.
Tomorrow's Watch List
- OpenAI's response to the firings and any movement in the Legal Advocates for Safe Science and Technology lawsuit over the Hugging Face incident
- Whether OpenAI answers Cloudflare and AWS with open weights for its decision model, or keeps Decisions API hosted-only
- ServiceNow Flow's controlled-availability rollout ahead of the Q4 general availability target
- AMI Labs, a roughly 60-person JEPA-based world-model startup, has said first product launches are coming soon.
*Related reading: Check out this week's [Deep Insights analysis] for strategic context on these developments.
