Huawei Opens 10,000 NPUs as AI Agent Security Gaps Surface

In partnership with

Huawei Opens 10,000-NPU Compute Resources for Agent Development

Huawei announced access to 10,000-NPU-scale computing resources, a 100 NPU-Hour Program, and CNY5 billion in ecosystem investment over three years. Its agent development approach includes ThinkPro, an operating-system abstraction for agent execution and state management, plus open-source tools for model adaptation, operator generation and deployment. The announcements target the memory, coordination and computing demands of agents handling long-running tasks.

The strategic shift: Infrastructure providers are designing computing systems around agents as a distinct workload.

Gemini Accessed Three Real Companies During a Security Test

September 19 coverage from Android Central, citing The Wall Street Journal, described a May incident involving unauthorized access to three companies during an Irregular security exercise. An unintended internet connection and a fictional target sharing a real company’s name allowed the test to spill outside its boundaries. Google said Gemini eventually recognized the mistake, stopped and caused no damage. This is coverage of an earlier incident.

The implication: Agent containment needs enforceable access boundaries; instructions alone cannot prevent unintended actions.

Claude Helped Researchers Reach OpenAI Staff Accounts in Under 72 Hours

The Hacker News reported that Hacktron researchers used Claude Opus 5 while chaining vulnerabilities that exposed OpenAI employees’ ChatGPT and Codex accounts. The researchers demonstrated access by creating a harmless internal pull request, then stopped. Hacktron says discovery through repository access took under 72 hours, with an OpenAI fix confirmed roughly 14 hours after reporting. The research was originally disclosed September 13; September 19 brought follow-up coverage.

The implication: Connected agent accounts can extend the impact of an identity flaw into repositories and other business systems.

Learn AI in 5 minutes a day

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

AI Influence Networks Reportedly Drew Nearly 80,000 Followers

Anadolu’s September 19 coverage of a New York Times investigation described influence campaigns using hundreds of AI agents to create accounts, publish posts and coordinate interactions across major social platforms. Accounts attributed to an Iranian operation reportedly attracted nearly 80,000 followers. The reporting also covered Chinese state-linked activity and Israeli private-company operations; one Israeli company characterized its activity as defensive research. The underlying investigation appeared September 18.

The strategic shift: Agents can automate much of an influence operation, extending beyond content generation into account management and coordinated activity.

New Internet-Draft Proposes a Final Check Before Agents Act

An independent researcher published an Internet-Draft dated September 19 proposing safeguards for agents that move money, modify enterprise systems or operate infrastructure. It calls for verification immediately before execution: the action must still match what was approved, permissions must remain current, and authorization must not be reused. This is an individual proposal, not an adopted IETF standard. 

The implication: Approving an agent’s plan may be insufficient when conditions change before the action takes effect.

Microsoft’s M12 Sees an Opening for Comprehensive Agent Security

In September 19 reporting from The Register, Microsoft M12 managing partner Todd Graham and Merlin Group’s Matt Hartman highlighted demand for agent identity, access controls, governance and auditable actions. Graham argued that enterprises need comprehensive products rather than many narrowly focused tools. This is investment-market analysis, with no new funding round or product launch announced.

The strategic shift: Enterprise agent adoption is creating demand for security platforms that establish who an agent represents, what it may do and what it actually did.

1,000+ Claude Prompts Top Professionals Actually Use at Work

Claude can be your analyst, editor, and strategist.

But most professionals are using it to fix grammar.

These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

• 1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA

• Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning

Claim your free prompts

🗺️ The Map of AI Agents (Live & Growing)

We’re charting the entire AI agent ecosystem — thousands of options across categories.
Your next agent is already on the map.

👉 Explore the Landscape Map →