Cybersecurity

Google DeepMind has pulled off something cryptographers have talked about for years. On November 10, the lab said its AI agents mathematically proved that the code behind Signal’s encryption does what it promises, finishing in four weeks. The work covers the protocol guarding private chats in Signal, WhatsApp, and Google Messages, touching billions of conversations a day. The project is called AlphaProof Code. A five-person team guided Gemini Carbon agents as they generated more than six million lines of proof in Lean, a language built for checking mathematical arguments. Lean’s kernel then checked every line, and a proof that doesn’t hold gets thrown out. ...

Google opened Gemini 4 Carbon, its most powerful AI model, to paying developers and businesses on Tuesday, November 10. The model had been locked behind a limited cybersecurity program since late October. Now anyone on the paid Gemini API tier can call it. The price is $2 per million input tokens and $10 per million output tokens through November 30. On December 1, both rates double, to $4 and $20. Google AI Ultra subscribers also get Carbon in the Gemini app, AI Mode in Search, and the Antigravity coding platform. ...

Fireworks AI says hackers accessed customer billing data, datasets and unfinished model weights in two breaches — and the company is still investigating the full extent. Full story in slides above — Sources: Forbes, The Neuron, Fireworks AI security bulletin, November 9-10, 2026.

A single click in Anthropic’s Claude Code has been quietly copying developers’ passwords and API keys into a file that anyone on GitHub can read. Security firm Apiiro said on November 5 that 1,122 public copies of that file hold at least one secret, and the total reaches 2,539 secrets across 1,994 repositories once git history and shared settings files are counted. The cause isn’t a hack or a prompt-injection trick. It’s a convenience feature doing exactly what it was built to do. ...

Google DeepMind announced on November 6 that its AI agents, guided by human cryptographers, uncovered two weaknesses in legacy 1024-bit RSA encryption — one making digital signatures easier to forge, the other making it cheaper to factor many old keys at once. Modern 2048-bit keys are not affected, DeepMind says. Sources: Google DeepMind, CNBC, Data Studios, November 6, 2026.

Google DeepMind turned its AI agents loose on one of the oldest locks in internet security, and the agents came back with two new ways to pick it. The lock is 1024-bit RSA, an encryption standard the industry retired years ago. Your bank account is not in danger. DeepMind AI RSA encryption research was announced Friday, November 6, by Pushmeet Kohli, DeepMind’s vice president of science, in a post on X. Two technical papers back it up. The work was led by Dj Dvijotham and Abhradeep Thakurta, working with outside cryptographers Nadia Heninger of UC San Diego and Inria and Emmanuel Thomé of Inria. Everything ran in a sandbox, against a retired standard chosen precisely because attacking it harms nobody. ...

OpenAI released its most capable hacking model yet on Thursday, and almost nobody is allowed to use it. GPT-6 Cyber, a cyber-specialized version of GPT-6 Astra, scored 48.5% on the company’s ExploitGym benchmark, a test that asks a model to turn a known flaw into a working exploit against a remote target. That’s up from 32.1% for the previous model, GPT-5.6 Cyber, and 42.4% for Astra itself. OpenAI rates the new model “Critical” for cybersecurity, the top tier of its Preparedness Framework, and it is handing it out only to a select group of vetted customers. ...

OpenAI says an autonomous pipeline built on an upcoming cyber model found, exploited, and drafted patches for 948 distinct Linux kernel privilege-escalation bugs in a single mid-October run, at a cost of $1,213,430 in API credits. Full story in slides above — Sources: OpenAI, “A thousand fewer paths to root” (November 4, 2026); GBHackers; StartupFortune.

OpenAI’s security team set its AI loose on the Linux kernel and came back with a number that sounds made up: 948. In a single 10-day run in mid-October 2026, an autonomous pipeline built on an upcoming cyber model found, reproduced and exploited 948 distinct local privilege-escalation vulnerabilities in the Linux kernel, bugs that let an ordinary user account take over a machine as root. OpenAI published the results Wednesday, November 4, in a research post titled “A thousand fewer paths to root.” The whole run cost $1,213,430 in API credits, or about $1,279 per working exploit. ...

Chainalysis, the blockchain analytics firm that police departments, banks and crypto exchanges use to trace stolen digital money, disclosed on October 30, 2026 that its own systems had been breached. An outside attacker, the company said, used “offensive agentic techniques” to get into its systems and internal tools earlier in October and pull out customer identity data. The Chainalysis breach exposed four fields: user IDs, full names, email addresses and organization role labels. Chainalysis called that data non-sensitive, a label that will do little to calm the people on the list. A name, a work email and a job title tell an attacker exactly who handles investigations or compliance at a given company, and where to aim a phishing email. ...

Polymarket’s international exchange lost about $2.7 million on Thursday, October 29, 2026, after a hijacked page-speed script turned logged-in traders’ own browser sessions against them. That’s the short version of the Polymarket DebugBear hack, and it’s worth reading slowly, because no smart contract broke and no password was cracked. Reports published November 1 put the affected count at 726 wallets and total refunds at about $3.9 million, a figure that includes $2.7 million returned to drained wallets plus about $1.19 million for roughly 5,500 wallets hurt by the trading pause. Polymarket halted trading for about an hour on October 29, then brought the international platform back online the same afternoon. ...

For about 11 hours on Thursday, October 29, Cursor’s extension marketplace served its users an update to Anthropic’s Claude Code extension that Anthropic didn’t write. Version 2.1.313 arrived under the extension’s real publisher ID, anthropic.claude-code, so anyone with auto-updates on could pick it up without clicking anything suspicious. Cursor says at most 5,000 users received the tampered build. That’s a small number next to Cursor’s user base. It’s a big deal anyway, because of where the package came from and what it tried to do. ...

Binance says hackers drained about $6.3 billion from its wallets on Thursday, in what investigators now call the largest crypto theft ever recorded. The world’s biggest crypto exchange put the estimated impact at $6.3 billion, while blockchain firm TRM Labs traced $6.36 billion moving out of Binance wallets across 10 blockchains. Most of the money moved in about 20 minutes, TRM said, starting at 20:05 UTC on October 29. Binance has since worked with partners to freeze roughly $3 billion of the stolen funds, and the company says it will cover all losses. Withdrawals remain paused while Binance overhauls its withdrawal systems, with service expected back within 48 hours. ...

Mistral Large 4 open weights landed on October 29, 2026, and they change who gets to run a frontier-class model. France’s Mistral AI published the full weights for Large 4, nicknamed Le Chonk, on Hugging Face three weeks after its October 6 preview. Anyone can now download it and run it on their own servers. Don’t expect to run it on a laptop. Large 4 holds about 1 trillion parameters in total, with roughly 49 billion active for each token, and it takes text and images with a 1-million-token context window. Through Mistral’s own API, it now costs $0.70 per million input tokens and $2.10 per million output tokens, about half the preview price of $1.36 and $4.18. ...

Google unveiled Gemini 4 Carbon on October 28, 2026, its newest frontier AI model built for cybersecurity and long engineering tasks, with introductory pricing of $2 per million input tokens and $10 per million output tokens. Full story in slides above — Sources: Reuters, SiliconANGLE, ET CISO, October 28-29, 2026.

Moonshot AI just did something almost no AI lab volunteers to do: it made its newest model worse at one job, on purpose. On Saturday, the Beijing-based company said it had stripped cyber-related data from the training of Kimi K3.1 Pro, its most capable open-weight model, while keeping the general coding skills that made the Kimi line popular with developers. Reuters reported the curbs as Moonshot introduced the model on October 23, and the company plans to publish the model’s weights by November 4. ...

AI researchers disclosed this week that AI models helped find PHP2Shell, two critical flaws (CVE-2026-109618 and CVE-2026-109619) in PHP, the programming language behind roughly 70% of the world’s websites, including WordPress and Drupal. Wiz says it found the memory-safety bugs using its Wiz Atlas agent with Gemini 4 Argon and GPT-5.6 Sol, while Calif independently found one flaw with GPT-5.6 Sol through OpenAI’s Patch the Planet programme. Patches released on October 21 cover every PHP version back to 5.4, but a Censys scan showed about 1.5 million of 2.6 million examined hosts were still unpatched, and Amazon and Cloudflare reported attack attempts already climbing into the millions. ...

On Wednesday, PHP’s security team did something it almost never does. It shipped emergency patches for two critical flaws (CVE-2026-109618 and CVE-2026-109619, collectively nicknamed PHP2Shell) that let an attacker run code on a web server without logging in first. The bugs are rated 9.2 out of 10 on the CVSS v4 severity scale, sit in the PHP language itself rather than any application, and one of them was found with the help of OpenAI’s models. ...

One hacker broke into more than 100 company websites. He barely lifted a finger. The AI agents did almost everything. Over the past couple of weeks, a Chinese-speaking hacker compromised websites belonging to more than 100 companies, including farm equipment maker Deere, appliance company SharkNinja and engine manufacturer Cummins, according to an investigation Reuters published on October 19. Each successful break-in cost the attacker about $50 in AI tokens. The human’s role in the whole operation? He wrote fewer than 100 Chinese characters of instructions, then walked away. ...

The Department of Homeland Security has ordered all federal civilian agencies to remove DeepSeek AI products from their information technology systems and barred future purchases, per a Federal Register notice published Monday. Full story in slides above — Sources: Wall Street Journal, Firstpost, Economic Times, October 19, 2026.