we_are_coded.by CODE · The world, decoded
БГ
Analysis

In one day, agents got their own processor, their own identity, and their own attackers

NVIDIAInfra

On 7 July 2026, three unconnected announcements landed on the same day: NVIDIA unveiled a processor designed around agentic workloads; Vercel bought Better Auth to build identity for agents; and CISA confirmed active exploitation of Langflow - a platform agentic workflows run on. Infrastructure, identity, attack.

In short
  • NVIDIA Vera: an 88-core CPU pitched as a processor for agentic workloads; Perplexity measures 1.5x faster task completion.
  • Vercel buys Better Auth (stays MIT, free) to build identity for agents - their own, scoped, revocable.
  • CISA confirms active exploitation of Langflow: someone else's agentic flow runs with someone else's flow ID.
Checked on8 July 2026Responsible editorTsvetelin IvanovHow we workMethod · Corrections

Three companies. Three occasions. Zero coordination between them. One and the same day - 7 July.

The facts: (1) NVIDIA announced Vera - a CPU with 88 cores and 3.4TB/s core-to-core bandwidth, pitched explicitly as a processor for agentic workloads; Perplexity reports 1.5x faster task completion and 1.9x faster startup for concurrent sandbox-instances compared to x86 (the usual processors until now), and plans production deployment. (2) Vercel acquired Better Auth - a TypeScript authentication library with over 4.7 million weekly downloads; it stays free under the MIT license, and the team will work on agent identity and the Agent Auth Protocol, where every agent carries its own, scoped, and revocable authority. (3) CISA added CVE-2026-55255 to the Known Exploited Vulnerabilities catalog: an authorization bypass in Langflow (a platform for agentic automations) lets an authenticated attacker execute someone else's flow by supplying their flow ID. Primary sources: blogs.nvidia.com, vercel.com, cisa.gov.

I'm looking at the system, not the three news items separately. Each one on its own is a technical detail. Put together on one day, they say something else: a whole ecosystem is already assembling around agents.

A technology stops being a toy the day it gets hardware under it, identity around it, and an attacker against it.

This is a familiar life cycle. Web servers went through it. Databases went through it. Containers went through it. First they're a tool, then someone casts silicon specifically for them, then comes the question of who has the right to what, and finally the people who turn it against you show up. Agents went through it in one day - not because anyone announced a stage, but because three independent companies hit the same pain point at the same time.

Langflow is the quietest of the three and the most important. The vulnerability isn't exotic: an authenticated user submits someone else's flow ID and the flow runs. A classic access rights problem. Except the object being executed is no longer a query to a database. It's an automation, with its own keys, its own access, and its own outbound calls. Permissions stopped guarding data. They guard actions.

That's why Vercel's move isn't a corporate footnote. Once an agent acts, it has to carry an identity - and it has to be revocable. Not inherit your admin key. Not live in an environment variable. Its own, scoped, pullable back.

That's why I'm looking at our own flows too. Every agent that runs a command, uploads a file, or touches production moves under someone's identity. If it's yours, you've handed the automation a human's rights. The debt comes due the day someone finds your Langflow. And it's already been found - at other people's places, this week.

The visual is generated code art. No third-party images.
Follow usFacebookLinkedIn
Official primary sources
→NVIDIA - Vera: max single-threaded CPU at scale→Vercel acquires Better Auth→CISA - Adds Three Known Exploited Vulnerabilities to Catalog
Original: https://wearecoded.com/en/articles/agents-get-silicon-identity-attackers.html
ShareFacebookXLinkedInTelegramWhatsApp
← Back to all news