Daily Agent News, Sunday, 1 November 2026

Good afternoon. Today's lead: Chainalysis, the firm police and governments use to trace crypto crime, says an attacker used "offensive agentic techniques" to break into its internal tools, and Wiz says at least nine AI developer platforms were breached in October as attackers hunted for AI keys to resell. Also today: Android Studio opens up to more than 40 outside coding agents, Perplexity Computer starts driving your own signed-in Chrome, AI agents now steer a radio telescope hunting for alien signals, OpenAI will make ChatGPT and Codex share one usage allowance from December, and a restaurant agent that lives in your texts.

LEAD: Agents are now doing the breaking in, from Chainalysis to "Black October"

Background: For months, security researchers have warned that AI agents would make attacks cheaper and faster. Earlier this week Microsoft said North Korean groups had taken up the open-source ARTEX attack agent. AI developer platforms are a juicy target because they store customers' API keys for OpenAI, Anthropic, Google and Amazon Bedrock.

What happened: On 30 October, Chainalysis published a notice saying that "this month" it detected an external actor "using offensive agentic techniques to gain unauthorized access to Chainalysis systems and internal tools." The attacker reached user IDs, full names, email addresses and organisation role labels for users on its platform. Chainalysis says it has found no evidence that investigation content (graphs, notes, cases), query results, clear-text passwords, MFA secrets or session tokens were accessed, and that the incident is not related to the recent theft at Binance. It hasn't said which vulnerabilities were used, how long the attacker was inside, or what "agentic" meant here, and it declined The Block's follow-up questions. It is now using large language models to help with a wider security review.

The same day, Wiz published "Black October", listing at least nine AI developer platforms that disclosed incidents this month: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank. Fireworks, OpenCode and Warp reported signs of an "agent swarm" attack, and every public report describes attackers hunting stored environment variables, especially AI keys. Wiz followed two stolen enterprise Bedrock keys: one made about 800,000 calls in 30 hours, the other about 1.7 million in 13 hours, almost all to Claude models, led by Opus (Wiz says three quarters of the first key's successful calls went to Opus), with traffic that looked like resold coding-agent sessions. This is "LLMjacking": running up someone else's model bill and selling the access cheaply. Wiz says Zeabur will shut down because of its breach. Northflank, the latest, told customers on 30 October to rotate every secret after users reported unauthorised use of their AI keys.

Why it matters: A security company with a mature team still had internal tools reached by an agent-assisted attacker, and the names and job roles it lost are exactly what targeted phishing needs. For anyone running agents, the practical lesson is to revoke and reissue AI keys at the provider whenever a platform reports a breach, and to set hard spend limits.

Sources: Chainalysis notice · BleepingComputer: Chainalysis · The Block · Wiz: Black October · BleepingComputer: Northflank

Android Studio lets you bring any of 40+ coding agents

Background: IDE makers have mostly pushed their own built-in assistant. The Agent Client Protocol (ACP) is an open standard that lets any compatible agent plug into an editor.

What happened: Google made Android Studio Rabbit 2 stable with "Bring Your Own Agent". Developers can pick from more than 40 ACP agents, including Anthropic's Claude Agent, OpenAI's Codex, JetBrains' Junie and Google's Antigravity, and those agents get access to the Android Emulator, build systems, Logcat and Android "skills". It also adds a skills manager and an AI Change Review that groups an agent's edits into themes for a guided walkthrough.

Why it matters: The world's main Android tool now treats the agent as swappable, so teams can keep their existing model contracts, and agents can test their own changes on an emulator rather than handing back untested code.

Source: Android Developers Blog

Perplexity Computer now drives the Chrome you're already signed into

Background: Perplexity Computer is the company's desktop agent for Mac and Windows. Until now, full browser control was built into Perplexity's own Comet browser.

What happened: Perplexity said on 30 October that a new extension lets Computer work inside Chrome, Edge, Brave, Arc and other Chromium browsers, using sites where you're already logged in. The Chrome Web Store listing says it can open sites, read pages, fill forms and finish tasks across tabs, such as pulling invoices from a billing dashboard. Perplexity's demo turned a business's reviews into a prioritised list of fixes and replies. The listing says it only uses the browser during a task you started.

Why it matters: Driving a real signed-in browser is where agents become genuinely useful, and also where prompt injection from a malicious page is most dangerous. Perplexity hasn't said how it defends against that.

Source: Creators Toolbox write-up, citing Perplexity's post and the Chrome Web Store listing

AI agents now steer a telescope searching for alien signals

What happened: NVIDIA says Breakthrough Listen and the SETI Institute have put agents in charge of day-to-day work at the Allen Telescope Array in Northern California. From one plain-English request, agents can steer the antennas, calibrate the instrument, catch and fix data errors, and run AI models on live data. An orchestrator called Nexus acts as the agent harness, and the system handles 2.4 terabits a second on NVIDIA RTX PRO 6000 Blackwell GPUs. The pipeline has already made what NVIDIA calls the first real-time AI detection of a pulsar, in the Crab Nebula. The team is testing open models for observatories that might one day run offline on the far side of the Moon.

Why it matters: It's a vivid example of agents operating physical scientific instruments, with work that took humans weeks compressed into hours. To be clear, the pulsar is a natural object, not an alien signal.

Source: NVIDIA Blog

OpenAI will make ChatGPT and Codex share one token allowance from December

What happened: An OpenAI Help Center page updated in the last day says that from December, Plus, Pro and Business plans will share one usage allowance across ChatGPT and Codex, "as we bring the agentic capabilities of Work into the main ChatGPT experience." Usage will be measured in tokens, so bigger models and harder tasks use more; Instant chat won't count. OpenAI expects most users won't hit resets more often than now, and users will be able to buy credits once they run out. No exact date or allowance sizes yet.

Why it matters: It's another sign OpenAI is folding its agent into everyday ChatGPT, and heavy Codex users will now be drawing from the same pot as their chats.

Source: OpenAI Help Center

Zest puts a restaurant-booking agent in your text messages

What happened: Zest, a restaurant discovery app backed by Alexis Ohanian's 776, launched a concierge you simply text, at a New York number. It learns your taste from your actual card transactions (food and drink only, via Plaid), checks availability and books through OpenTable, Resy and SevenRooms. Drop it into a group chat and it blends everyone's tastes, runs a vote in the thread, then books for the group. TechCrunch found it "worked fairly well" in testing. "The agent is the product now," says founder Mario Gomez-Hall.

Why it matters: It's a neat example of agents turning a whole app into a chat thread, and of personal data, here spending history, being the edge over general assistants.

Source: TechCrunch

Also worth a look

• Dataline raises $10.6M for a crypto-based marketplace where agents pay per call for data, each answer carrying a confidence score; co-led by Lattice and Dispersion Capital, with Coinbase Ventures among backers. Its usage figures are its own claims. GlobeNewswire
• Omnichat raises a $10M Series A led by Openspace Capital for agents that market, sell and serve customers inside WhatsApp, LINE, WeChat and other chat apps; it says about 300 brands took up its OmniClaw agent within two months of launch. VC Update, citing PR Newswire