DeepSeek Unveils a Super-Agent Design as Wiz Tallies Nine AI Platform Breaches and an Open Model Exploits Firefox’s Patch Gap

DeepSeek Publishes a Super-Agent Design for Agents That Stay Online and Work in Parallel

DeepSeek used Saturday to argue that the next limit on AI agents is not the model but the person supervising it. In an article published by its DeepSeek Harness team, the company proposed what it calls a “Super-agent”: a single, always-on agent that keeps context across chat and code review, runs several jobs at once and steps in on its own when something needs doing. The Global Times report on DeepSeek’s Super-agent design says the team found that one person can keep track of only three to five active agent sessions at a time, and that most agents still force users to switch among separate sessions and manage each one’s context by hand.

The design splits the work into two layers. A resident router handles communication with people, keeps the shared context and hands out tasks, while multiple workers run assignments in isolated sessions and report progress back to the router rather than replying to users directly. The system is driven by events, such as a chat message, a code review comment or a CI result, and an external monitor called Watchdog checks for failures and recovers from interruptions. According to the Global Times, DeepSeek said an internal Super-agent is already running, reproducing and fixing software problems, responding to code reviews and improving DeepSeek Harness itself, sometimes spotting a bug discussed in a group chat without being asked.

DeepSeek also released code. The dsh-superagent-prototype repository on DeepSeek’s GitHub describes an MIT-licensed, proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu, opens a work item when an allowed person asks for something, works on it in its own session and directory, and answers in the original conversation, for example with a pull request. The README is blunt that the published code differs from DeepSeek’s internal version and that some ideas in the article are future directions rather than features.

“Research prototype for reference only, not an official DeepSeek product.” — DeepSeek, dsh-superagent-prototype README

Wiz Counts Nine AI Developer Platform Breaches in “Black October” as Northflank Tells Customers to Rotate Every Key

Security firm Wiz has tied a month of separate incidents into one picture. Wiz’s “Black October” report on attacks against AI developer platforms lists at least nine platforms that disclosed incidents in October: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank. Wiz says every public report describes attackers hunting stored environment variables, with a focus on those that grant access to AI services. It found no single vulnerability or dependency linking the attacks, but says the pattern fits two trends: agentic attacks and LLMjacking, in which stolen AI keys are used or resold. OpenCode and CommandCode attributed their incidents to attempts at distillation, and Zeabur is shutting down after its breach.

Wiz traced two enterprise keys that had been heavily used from a compromised vendor’s infrastructure. Both were first tested within 30 minutes of the reported credential access. In one case, an attacker made about 800,000 Amazon Bedrock calls over thirty hours, roughly 90% of them successful; in the other, a single address made about 1.7 million calls over thirteen hours. Wiz estimates the two cases alone likely ran up more than $100,000 at list price, and it linked the keys to infrastructure tied to two discounted model-resale sites, while saying it cannot prove those operators carried out the original theft.

The latest victim disclosed on Friday. BleepingComputer’s report on the Northflank breach says the cloud hosting platform received a threat alert from Wiz at 11:43 UTC on October 30 (4:43 AM PDT), patched its network policy and later confirmed that customers had reported unauthorized use of AI API keys stored on the platform. Northflank has not said how attackers got in or how many customers were affected.

“Assume all environment variables have been decrypted. Disable/rotate all keys,” — Northflank, via BleepingComputer

Epoch AI Shows an Open-Weight Model Exploiting Firefox’s Patch Gap for About $1,000 a Version

A widely available open model can now turn published-but-unshipped security fixes into working browser attacks, according to research from Epoch AI. Epoch AI’s study on open-weight models exploiting the patch gap found that GLM 5.3, released by Z.ai in August, built attacks against all six Firefox releases between September 17 and October 26 using only information available when each version shipped. Visiting the webpage it produced made the browser read a local file and send it to the attacker with no further user action. The patch gap is the window between a fix appearing in public code and reaching users.

Epoch says the model worked autonomously after basic instructions, picked out security fixes that Firefox disguises as ordinary improvements and chained them into multi-stage exploits. Its fastest attempt took 12 hours, and once the harness was built, the marginal cost of compromising a new version was about $1,000 at API prices. Every successful attack on the latest versions also relied on an unpublished vulnerability the model rediscovered, which Mozilla patched in Firefox 159.0 on October 27. The same setup has not yet produced an equivalent exploit for Google Chrome, despite about $100,000 spent on attempts. The Decoder’s summary of the Epoch findings notes the researchers framed the work as a security exercise.

“We did not trigger any safeguards or refusals from any provider, including Z.ai, when using GLM 5.3 in this way.” — Epoch AI

Pope Leo XIV Says AI Systems Are “Tools, Not Persons” and Urges a Ban on Lethal Autonomous Weapons

Pope Leo XIV used an audience on Friday to lay out five principles for AI. Speaking to participants in a meeting organized by the Pontifical Academy for Life and the Future of Life Institute, the pope said the future of AI is not predetermined and that the key question is not what AI can do but what purpose it should serve. The Holy See Press Office text of the pope’s address on AI shows him rejecting proposals to give AI systems rights or legal personhood, arguing that machines are tools, not persons.

He said decisions affecting human life, freedom and the use of force must always remain subject to effective, informed and responsible human oversight, and encouraged the states meeting in Geneva in November to take concrete steps toward banning lethal autonomous weapons. His final point was institutional: the ethical principles are largely known, he said, and what is missing are effective international bodies with shared standards, independent oversight and clear accountability. Vatican News coverage of the address framed it around keeping the human person at the heart of the technology.

“Such an approach risks obscuring human responsibility and leaving those who suffer harm to confront a faceless entity.” — Pope Leo XIV

Intel Posts Its Worst Day Since the 1970s as a Memory Crunch Hits PCs and AI Servers Keep Growing

The AI boom’s side effects reached Intel’s stock on Friday. CNBC’s report on Friday’s Intel and Amazon selloff says Intel shares fell 29% on Friday, the stock’s worst day since the 1970s, after the company guided fourth-quarter revenue to $15 billion to $16 billion against analyst expectations of $17.1 billion, according to LSEG. CNBC says the PC market is being hammered by soaring memory and component costs, and that an explosion at Intel’s fab in Israel will cost about $600 million in lost revenue over two quarters. The drop cut the value of the U.S. government’s Intel stake by about $12 billion, and the Nasdaq Composite slid 2.9% on the day.

The data center side of the business looks very different. Intel’s third-quarter 2026 earnings release shows total revenue of $16.5 billion, up 21% year over year, with its Data Center and AI group up 78% to $7.3 billion while its client computing group fell 5% to $8.1 billion. Chief executive Lip-Bu Tan said demand for Intel’s CPUs remains robust and that the immediate priority is unlocking supply to meet what he called extraordinary demand for Xeon. The AI link runs through memory: in CNBC’s roundup of analyst reactions to Intel’s guidance, UBS analyst Timothy Arcuri said he sees no relief in sight for memory pricing and supply, especially given how much server demand is growing for agentic applications, and Bank of America downgraded the stock to neutral.

“Looking ahead, we see the PC slowdown continuing at least through the first half of next year” — David Zinsner, Intel CFO

Google DeepMind’s PlanAssist Expands to 15 English Councils After Cutting Assessment Time by 58%

A Gemini-based planning tool is moving from pilot to a wider live beta in England. Google DeepMind’s post on speeding up planning decisions with PlanAssist says the Gemini-based tool cut the average time officers in Barnet, Camden and Dorset spent on householder assessment tasks by 58% in early pilots, and that more than 70% of the reports it generated were approved by officers with no edits. PlanAssist validates documents and drawings, checks applications against heritage constraints and local and national policies, looks for prior objections and drafts cited reports that officers can edit before deciding.

The UK government’s press release on faster planning decisions says the tool, developed with Google DeepMind, Google Cloud, Faculty and local authorities, will expand to 15 councils for 12 months, with the government intending to make it available to every council in England by the end of 2027. The 12 new councils are Ashford, Lambeth, Newcastle, Nottingham, Bristol, Cumberland, North Lincolnshire, Rochdale, Redbridge, South Gloucestershire, Stoke-on-Trent and Westminster. Officers remain responsible for decisions, and the software will not be mandated.

“This is what putting AI to work for the public looks like in practice.” — Kanishka Narayan, UK AI Minister

NVIDIA and the SETI Institute Put AI Agents in Charge of the Allen Telescope Array

The search for extraterrestrial intelligence has gained an agent harness. NVIDIA’s blog post on AI joining the search for extraterrestrial signals says Breakthrough Listen and the SETI Institute now let agents at the Allen Telescope Array in Northern California steer the instrument, correct incoming data and apply AI models to the live stream from a single natural-language request. NVIDIA’s Holoscan platform pulls sensor data straight into GPU memory, the system handles 2.4 terabits per second on RTX PRO 6000 Blackwell GPUs with inference running on premises, and the pipeline has already achieved the first real-time AI detection of a pulsar in the Crab Nebula. The team is now evaluating open models including Nemotron for observatories that may someday run offline.

“But this isn’t something to fear. Rather, it’s a breakthrough in translation.” — Andrew Siemion, SETI Institute and Breakthrough Listen