News at a Glance
OpenAI isn't a public supporter of Nvidia's Open Agent Safety Platform, but it is privately working with Nvidia, TechCrunch has learned.
What Happened
Nvidia unveiled the Open Agent Safety Platform at its autumn GTC in San Jose on September 29, with Jensen Huang calling it "a seatbelt for the agent era." The platform has three parts: an open identity specification requiring every autonomous agent to carry verifiable credentials that record its principal, permission scope and full call chain; a runtime policy engine that approves high-risk actions such as file writes, payments and external API calls one by one; and a circuit-breaker module that freezes an agent and rolls back its state within 200 milliseconds when it detects privilege escalation or cascading errors. The specification is released under Apache 2.0. The fir
Background
Over the past 18 months enterprise agents have moved from demos into production, calling APIs, moving money, editing code and sending mail on their own — and incidents followed: unauthorised access to internal databases, runaway loops that produced large bills, and production configuration changes made without human review. Earlier safety work focused on the model layer: Nvidia's NeMo Guardrails in 2023, MLCommons AI safety benchmarks in 2024 and the Frontier Model Forum's risk-disclosure frameworks. All of them addressed content filtering and model eval
Deep Dive
Liu Gong reads the missing signature as a fight over who defines the agent safety layer, not a fight about safety itself. From a technical standpoint, OASP turns identity credentials, policy enforcement and circuit-breaking into an open specification. Once mainstream clouds and security vendors adopt it, safety stops being a differentiator and becomes plumbing, much as TCP/IP became plumbing for the internet. The sharper move is binding credential verification to Nvidia's GPU secure-boot chain: an enterprise that wants a hardware-level root of trust has to run on Nvidia's compute stack. That is the CUDA playbook replayed one layer up. OpenAI's hesitation is easy to trace — its own agent sandbox and audit pipeline already ship inside ChatGPT Enterprise and AgentKit, so adopting OASP wholesale would hand years of safety engineering to a supplier and leave an externally auditable dependency list. On the business side, compliance is not a cost centre but a toll booth. OpenAI has spent two years pushing agents into checkout flows, enterprise workflows and code automation, where a single privilege-escalation incident triggers payouts and trust damage. Owning the policy engine and the log store is what lets it charge merchants for the Agentic Commerce rail. Signing on publicly would place Nvidia between OpenAI and the customers who pay it, weakening its control over pricing in the authorisation chain. For the industry as a whole, a vendor-led safety standard carries real value and real structural risk. OASP gives regulators a single audit vocabulary that fragmented in-house builds never could, which is why agencies are likely to lean on it. But when the specification is owned by a chip vendor, model labs lose bargaining room and user permission data concentrates at the compute layer. OpenAI's quiet engagement is a familiar rule-shaping tactic: shape the draft without endorsing a rival's framing. Liu Gong expects two OASP tracks to emerge within a year — a neutral public specification and an Nvidia-hardware-tied edition — with OpenAI shipping a credential-compatible but separately governed policy engine, forcing buyers to choose between an open baseline and a vertical closed loop.
Data Points
- First public backers: 24 vendors across model labs (Google DeepMind, Anthropic, Meta), cloud (Microsoft, Amazon), security (CrowdStrike, Palo Alto Networks) and devices; OpenAI and xAI are absent.
- Nvidia's stated benchmark: freezing a rogue agent and rolling back state within 200 milliseconds, versus 3 to 8 seconds for the average in-house audit pipeline from alert to human intervention.
- Regulatory exposure: EU AI Act fines for general-purpose model providers reach 3% of global revenue or EUR 15 million, whichever is higher, with traceable logs a p
Impact & Outlook
For the industry, the near-term effect is to move agent safety from bespoke in-house builds to a comparable public baseline. Within three months, procurement teams will likely write "OASP-compatible" into tenders, security vendors will ship adapters quickly, and OpenAI's silence will be pressed repeatedly on earnings calls. Within a year, if OASP becomes a de facto standard, competition shifts more purely to model capability while safety engineering drops from differentiator to entry ticket — a gain for smaller labs that can plug in and comply, but a cos
Angles to Explore
Points to consider
- Where an open specification ends and commercial lock-in begins, given that Nvidia owns both the compute layer and the secure-boot chain.
- How agent identity attaches to payments: who issues the checkout authorisation in agentic commerce, and who collects the toll.
- The cost of absence: can regulators treat OASP as authoritative when the largest consumer agent platform declines to sign.
Source & Original
This item comes from TechCrunch AI (published 2026-09-29 18:35:00). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.