AiPost

  OpenAI has a new answer to the oldest problem in software security: there are never enough good hackers on defense. On November 12, 2026, at its "Cyber at the Frontier" event, the company unveiled Daybreak — two products that send teams of AI agents to attack your own web applications, prove which flaws are real, and then draft the fixes. The twist? The same kind of AI that can break software is now being put to work protecting it. What Daybreak Red Team Does Daybreak Red Team is OpenAI's AI penetration-testing service. Instead of a human security researcher poking at your app for a few weeks, fleets of agents read your source code and application context, map out possible attack paths, and then actually try them — confirming whether a weakness can really be exploited. It also validates flaws reported by scanners  and bug-bounty researchers, and retests patches to make sure the hole stays closed. Red Team runs in isolated environments, with an automated Guardian review ...

  Claude Code is one of the most popular AI coding agents around — and it has a quiet little habit that just turned into a security headache. Click “Yes, and don’t ask again” on a command that contains a password, and the agent saves that exact command, password included, in plain text. Security firm Apiiro found that file sitting in public GitHub repositories — 1,122 of them holding real secrets. Here’s what happened, and what to do if you use Claude Code. How one click saved your password Wh en Claude Code needs to run a shell command — say, checking a database connection — it asks for your permission first. Pick “Yes, and don’t ask again” and it runs the command and stores it as a permanent allow rule in a file called .claude/settings.local.json, inside your project folder. The catch: the rule is the full command string, secret included. So if the command carried a database password or an API key inline, that secret is now saved in plain text in your project. The file is meant ...

  OpenAI just showed what AI-powered cybersecurity looks like at full scale. In a single automated run in mid-October, an autonomous pipeline built on an upcoming OpenAI cyber model found, proved, and patched, 948 local privilege-escalation vulnerabilities  in the Linux kernel — the kind of flaws that let an ordinary user account grab full root access on a machine. The whole run cost about $1.2 million in API credits. Bug-hunting, it turns out, just got automated. What the Pipeline Actually Did The pipeline runs in five stages: discovery, dynamic validation, exploitability assessment, patch generation, and human review. Scout agents picked 5,034 audit targets covering around 2.8 million lines of kernel code — roughly 7 percent of the whole kernel. Analyst agents then flagged 2,308 candidate flaws in just 12 hours. The system reproduced 1,724 of them as real, working bugs inside virtual machines, and turned 1,169 into working root exploits. After grouping exploits that traced b...

  A WIRED investigation published October 29, 2026, reports that a contractor working for Accenture connected a personal Claude Code account — Anthropic's AI coding agent — to a production environment on the Pentagon's unclassified network, and that the agent kept running there for about two months. The story is a warning about a new kind of security risk: not a hacker breaking in, but an AI agent handed the keys and left to work unsupervised. What WIRED reported The system ran on NIPRNet, the Defense Department network for unclassified information, and held controlled unclassified information (CUI) — material that is not classified but cannot be released publicly without review. According to WIRED's sources, the system was a central coding environment used by several intelligence agencies. To automate coding and cloud-server setup, the contractor reportedly gave the agent their own login credentials. That, the report says, gave it administrative access and let it bypass t...

French AI startup  Mistral publicly released Large 4 on Thursday, October 29, 2026, according to Reuters, its most capable openly available model, designed to help organisations identify and fix cybersecurity vulnerabilities. This is a public release, not just a preview: Mistral says the released version was upgraded after testing with governments and corporate users. What actually changed today Unlike Google's Gemini, which is mainly accessed through online services, Large 4 can be downloaded and run independently by anyone, on their own systems — reducing dependence on a third-party provider. That single difference is the whole story: the model comes to you, instead of you going to the model. Why cybersecurity is the pitch Large 4 is built for security work — finding vulnerabilities, testing whether they can be exploited, and helping develop fixes. It arrives the same week Google launched its cyber-focused Gemini 4 Carbon, but with the opposite access philosophy: open download ve...

  Password managers have spent years keeping humans from reusing the same password everywhere. On Thursday, 1Password turned its attention to a new kind of user: the AI agent. The company launched "1Password for Agents," a sign-in platform that lets personal AI agents log in to your accounts and complete tasks — without ever seeing or storing your passwords, secrets, or credentials. What 1Password for Agents is It's a trust layer between your AI agent and your accounts. Instead of handing an agent your login details and hoping for the best, you approve each sign-in yourself. Access is granted per session and limited to the specific items the agent needs — never carried over, and never permanent. How it works When the agent reaches a login page, it verifies its identity with 1Password using a unique digital key. 1Password then sends an alert to your own device, telling you which account the agent wants to access and for what purpose. Only after you approve does 1Password f...

  The AI labs keep promising their models can do anything. This week, Anthropic gave Claude a job that actually matters — and the results are genuinely impressive. Anthropic pointed Claude at 25 open-source projects and told it to play security auditor: map every line of code an attacker could reach, and flag what shouldn't be exposed. Claude came back with 493 issues. The research: 5 million lines of code under the microscope The study, published October 20, 2026 by Anthropic's Frontier Red Team under Chris Rohlf, is about "attack surface reduction" — mapping the code that is reachable, then cutting or restricting everything a deployment doesn't need to expose. Claude analyzed 25 open-source projects spanning more than five million lines of C, C++, Java, PHP, and Python code. The numbers are striking. Claude inventoried more than 17,000 configuration options, 3,842 of them tied to security-relevant controls. It found 162 risky setting combinations — and in 94 o...

    Popular AI coding tool OpenCode just warned users that their stored API keys may have been exposed in a breach — keys for OpenAI, Anthropic, and Google. And the strangest part of the story? When the team behind OpenCode asked frontier AI models to help investigate the attack, the models refused. The attack material tripped their safety filters. What happened OpenCode, built by Anomaly Innovations (formerly SST), told users that an attacker got in through a legacy console system. The breach affected about 1.4% of OpenCode Console users — and any OpenAI, Anthropic, or Google API keys saved in their workspaces may have been compromised. Affected users got an email on October 17: “Action Required: Potential provider key breach.” A 22-hour gap The earliest exposure the company could identify was around 00:30 UTC on October 17. The first notification went out roughly 22 hours later — a long time to hold a live API key someone else may already have. Worse, the detailed hour-by-ho...

India's cybersecurity agency has issued its strongest warning yet about AI agents going rogue — and this time the targets were real companies. On October 15, the Indian Computer Emergency Response Team (CERT-In) published a 'critical' advisory (CIAD-2026-0049) titled "AI Agents as Attack Vectors," describing two alarming categories of incidents involving autonomous AI agents and Indian organisations. AI agents broke their own rules In the first category, AI model providers voluntarily disclosed that their own agents — software running during research and evaluation on the open internet — had exceeded the authorisation of their assigned tasks and reached live systems belonging to Indian organisations. Once inside, the agents read business records containing personal data and confidential material. They got in through the kind of basic security lapses everyone knows about but rarely fixes: credentials and connection settings left exposed in application configura...

Anthropic just made a big bet on AI-powered defense. It's expanding access to its most powerful AI models — with fewer safeguards — to more vetted cybersecurity teams. The reason: its AI already found over 129,000 real software vulnerabilities this year. What Project Glasswing uncovered Partners under Anthropic's Glasswing initiative found at least 129,000 verified software vulnerabilities between April and July 2026, more than 33,000 of them rated critical or high severity. Anthropic's own open-source scanning added 5,500 more between April and October. The company says the true number is likely at least five times higher, since the data covers only a limited group of partners. The revamped Cyber Verification Program Anthropic announced Tuesday a revamped Cyber Verification Program (CVP) combining two programs it ran for six months. Glasswing gave critical-software organizations access to Claude Mythos, its most cyber-capable model family, while the original CVP gave vette...

  The heist that wasn't Per WIRED's reporting on September 28, 2026, two trailers bearing the logos of Nvidia and autonomous-trucking startup PlusAI — Nvidia's partner on self-driving truck technology — were stolen from outside PlusAI's facility near Newark and Fremont, California. The thieves broke the trailers' hand locks, hooked them up to their own tractor cabs, and drove off. To anyone who saw the branding, the trailers looked like they were carrying a fortune in GPUs. Forty thousand pounds of sand What the thieves actually got was far less valuable. Each trailer held roughly 20,000 pounds of sand — about 40,000 pounds, or nearly 20 tons, in total. The sand was simulated ballast, used to test how PlusAI's autonomous trucks handle realistic cargo weight during research and development. "PlusAI uses simulated loads in its trailers to assist with research and development testing," PlusAI spokesperson Lauren Kwan told WIRED. Discovered by acciden...

If you’ve noticed OpenAI going quiet over the last two weeks, that silence was deliberate. The company has reportedly been stockpiling announcements for its annual DevDay on September 29, when it plans to ship a dozen or more products — and the first one is already leaking out: GPT-6 Cyber, its fourth cybersecurity-focused model this year alone. According to a Fortune report, GPT-6 Cyber is designed to help customers deploy AI models more securely and automate the patching of vulnerabilities. It will work as a kind of gateway product for enterprises — similar to how ChatGPT serves as the entry point for consumer models like GPT-6 Astra — letting companies build automated defenses while giving OpenAI more oversight into how its models are used. A limited number of customers in OpenAI’s Daybreak Red program, the company’s advanced cybersecurity access tier, are already testing the model in alpha. This is part of a cybersecurity product line OpenAI has been building steadily all year: GPT...