Trendora Radar — Volume 4
Volume 4 is led by a strong shift toward AI: headless AI agents moved from trial to adopt, while Artificial intelligence, web crawler, HTTP, Anthropic API, Zero Trust, localStorage, Chrome, NVIDIA GeForce RTX 3090, and F# all moved into trial. This week’s stories also center on AI and security, from DARPA and the U.S. Air Force testing AI-controlled F-16s and Anthropic’s copyright settlement to a firmware token exposure at Hanwha and a malicious Git hook hidden in a take-home project.
What moved on the radar
Top stories
- Hanwha camera firmware exposed a GitHub admin token
A security researcher analyzed Hanwha Vision camera firmware and recovered the root filesystem by reversing an obfuscated firmware upgrader. The extracted files contained a GitHub token with admin access to hundreds of repositories, likely leaked through build-time environment variables embedded into the camera UI files.
Hacker News - DARPA and U.S. Air Force test AI-controlled F-16s
A modified U.S. Air Force F-16 is undergoing human-on-the-loop flight tests with an AI agent controlling the aircraft under DARPA and Air Force programs. The VENOM Autonomy Kit lets pilots switch between human and AI control without changing the jet’s core software, supporting future live-flight testing of multiple AI agents and autonomous combat aircraft concepts.
Hacker News - Take-Home Interview Project Hid a Malicious Git Hook
A developer inspecting a take-home assignment discovered a pre-commit Git hook that would detect the host operating system and fetch a remote script for execution. The payloads were hosted on a raw IP address and appeared designed to run silently when Git operations were performed, indicating a malware delivery attempt disguised as a hiring exercise.
Hacker News - Anthropic copyright settlement approved
A judge approved Anthropic’s $1.5 billion settlement in a major copyright class-action case, resolving claims tied to the company’s use of books for AI training and alleged piracy. The deal is the largest copyright class-action settlement ever certified, with only 350 authors opting out.
Ars Technica AI - Hidden AI-related debts at US tech giants surge to $1.65T
A Nikkei study says off-balance-sheet liabilities at five major U.S. tech companies have risen sharply as AI spending expands, reaching an estimated $1.65 trillion. The report says data center leases and GPU supply contracts are making the companies' true obligations harder for investors to see, with Meta and Oracle highlighted as examples.
Hacker News - Romania's land registry hit by destructive breach
Romania's cadastre agency says a hacker wiped parts of its land registry systems after a failed extortion attempt, disrupting real-estate transactions and related public services for a week. The incident also involved stolen employee credentials, internal documents, and network details being offered for sale online, while officials say they are rebuilding the network from scratch.
Hacker News - Forecast Points to Record-Strong El Niño
Seasonal forecast models now give a roughly 90% chance that the 2026-27 El Niño will become the strongest on record, with a median peak Niño 3.4 anomaly of about 3.6°C. The forecast ensemble is unusually consistent and sits above the historical record set in 2015-16, though the article notes that model agreement does not guarantee forecast skill.
Hacker News - Relay market behind token resale and fraud
The article examines an underground relay market that proxies traffic to U.S. AI models at steep discounts, allegedly enabling token resale and widespread abuse. It describes a layered ecosystem of card merchants, account pools, and Chinese-language relay services that package access for downstream customers.
Simon Willison
Implications at publication
Headless AI agents have been promoted to Adopt, and the news shows the market is focusing on practical, safe, governed deployment. Cloudflare’s note on the web business model shifting toward AI agents, along with examples of agents for chat, Git hosting, and SRE operations, suggests commercial demand is taking shape.
BD teams will need to sell on operations and risk control more than on “smarter AI,” because buyers care about agents that can be deployed safely inside real organizations. Conversations will center on governance, integration with existing systems, and concrete value for support, SRE, or internal workflows.
Headless AI agents have moved from Trial to Adopt, showing a shift from experimentation to practical use. The supporting articles point to agents working toward goals, the need for governed deployment, and bottlenecks such as retrieval engineering, semantic search, and operating agents in real environments.
Developers will need to treat headless AI agents as software with their own lifecycle: prompt design, access control, logging, rollback, and failure handling become part of building rather than an experiment. Integrating retrieval, semantic search, and governed deployment patterns will take more time than wiring up a single API call.
The move from Trial to Adopt indicates headless AI agents are entering formal product plans. The news suggests agents are being used for real operations, deployed under governance, and treated as a way to reorganize work around objectives.
PMs will need to define products around goals and constraints rather than features, because headless agents are increasingly framed as shifting work from tasks to purpose. That means clearer priorities around policy, permissions, observability, and acceptance criteria based on outcomes instead of automation counts.
The radar has moved headless AI agents into Adopt, so testing is no longer an afterthought. News about runaway agents, governed agent deployment, and expectations that agents will take on operational work shows that safety and reliability testing are becoming default requirements.
QA and testing will have to expand from static output checks to verifying agent behavior: decision paths, when to stop, permission compliance, and runaway cases. This increases the need for simulation, monitoring, and regression tests across multi-step flows rather than only request/response tests.
Anthropic has been promoted from Trial to Adopt. The supporting articles show expanding commercial opportunity through enterprise platforms, but also barriers around access regions, government controls, and deployment costs that must be reflected in deals.
With Anthropic now in Adopt, BD can sell toward real deployments instead of just closing pilots, but needs clear terms around service regions, infrastructure integration, and product limits. The news on Claude GA on Microsoft Foundry and restricted access in some regions shows expansion opportunities come with tighter commercial and legal structuring.
Anthropic has been promoted from Trial to Adopt. The supporting news shows broader deployment readiness through Claude Sonnet 5 and Claude GA on Microsoft Foundry, but also signals regional limits, high compute costs, and uneven quality in some editing tools.
Claude is moving from trial to adoption, so developers will start wiring it into everyday workflows instead of treating it as a POC-only tool. At the same time, the Claude Sonnet 5 release, Claude GA on Microsoft Foundry, and reports of uneven behavior in Claude Code and edit tools mean devs need to verify compatibility, token efficiency, and regional constraints before putting it in production.
Get it in your inbox
A weekly digest of radar movement + top stories. One email a week, unsubscribe anytime.