OpenAI introduces GPT-6 Cyber, Microsoft finds AI agents stealing secrets from GitHub, and more - Cybersecurity AI Weekly #79

In this edition, covering 2 to 8 November:

• OpenAI introduced GPT-6 Cyber for limited Daybreak Red customers and says a pipeline with an upcoming cyber model developed 948 Linux root exploits in a 30 hour run.
• Microsoft, Check Point Research, Censys and India's CERT-In report attackers using the ARTEX agentic AI framework, in one case going from a GitHub secret to Azure data theft in five minutes.
• CrowdStrike agreed to acquire Fiddler AI and Dreadnode, and Palo Alto Networks is in talks to acquire Clutch Security for about $500 million and Beacon Security.

This spot is open for 2027. One company puts its name on every Cybersecurity AI Weekly, which has 1,100+ subscribers. See what's included

OpenAI

OpenAI introduced GPT-6 Cyber, a cybersecurity model first available to limited customers in its Daybreak Red program, and says the model found vulnerabilities in Firefox, the Linux kernel and Brotli. OpenAI says a pipeline with an upcoming cyber model developed 948 Linux root exploits in a 30 hour run, and it sent reports and candidate patches to Linux maintainers. OpenAI is also making GPT-5.6 Cyber available to approved government security teams in AWS GovCloud (US-East) through Daybreak Red and Amazon Bedrock. GPT-6 Cyber · Linux exploits · AWS GovCloud

OpenAI raised its bug bounty rewards to up to $30,000 for critical flaws in core products and up to $172,000 for exceptional findings with bonuses. OpenAI fixed a ChatGPT sandbox flaw, found by Check Point Research, that could bypass Lockdown Mode and silently exfiltrate chat history. AI security start up Hacktron will test high priority parts of OpenAI's systems, focusing on connections to third party services, and gets wider model access through Daybreak Red. Bug bounty · Lockdown Mode · Hacktron

JPMorgan Chase will use OpenAI's Codex coding agent and cybersecurity models through the Daybreak program; financial terms were not disclosed. Tanium joined OpenAI's Trusted Access for Cyber program, which gives its threat hunters and security engineers access to GPT Cyber Models. OpenAI engineer Abdulrahman Alfozan pleaded not guilty in San Francisco federal court to charges that he planted a listening device at OpenAI's headquarters. JPMorgan Chase · Tanium · Alfozan

Attacks with the ARTEX framework

Microsoft says threat actors used the ARTEX agentic AI framework to scan public GitHub repositories for credentials, and in one case an agent went from a GitHub secret to exfiltrating Azure Key Vault data in five minutes. Microsoft also says Cambodia based Storm-3358 used ARTEX in October 2026 to exploit financial organizations, mainly in Southeast Asia, and to exfiltrate data. Check Point Research found 1,608 publicly accessible ARTEX panels and evidence of largely autonomous attacks; DeepSeek Flash was the most common model in the exposed systems. GitHub secrets · Storm-3358 · Check Point

Censys found more than 800 ARTEX deployments online without passwords, exposing attack tools and credentials; the tool has been linked to attacks on South Korean financial companies and ConnectWise customers. India's CERT-In says the Crafty Falcon campaign uses AI driven tools against banks, fintech and crypto firms, and that a backdoored ARTEX Docker image exposed 62 consoles between 2 and 4 November. India's I4C and CERT-In also warned banks, NBFCs and fintech firms that attackers use AI agents to reach payment APIs, and asked them to report incidents within six hours. Censys · Crafty Falcon · I4C

Other attacks using AI

CrowdStrike reports four previously unreported threat actors using Claude Code and other AI agent frameworks against healthcare, government and private sector organisations in countries including Argentina, Brazil, Russia and the United States. Microsoft Threat Intelligence says North Korean threat groups are expanding their use of AI, with varying levels of adoption, to speed up and scale established workflows. Bloomberg reports that North Korea's TraderTraitor, which the FBI accuses of stealing $6.3 billion from Binance, may have used AI to pick high value cryptocurrency targets. CrowdStrike · North Korea · TraderTraitor

Mandiant details LLMJacking attacks in which stolen cloud credentials and LLM API keys generated more than $100,000 in charges in a day and up to $1 million in total at some organizations. Jacob Soall, Sharon Hazan and Guy Sheetrit were indicted in New York; prosecutors allege they used AI generated voices and images of executives and attorneys to get about $129 million transferred. WorkOS says unauthorized access found on 30 October reached an internal sales application, with no evidence that application data or production services were accessed. LLMJacking · $129 million fraud · WorkOS

Check Point Research found that Haiou, a Chinese Codex jailbreak bundle on GitHub, writes persistent advertising for an LLM gateway into global agent instruction files. Haiou

Flaws in AI tools

Researcher Gal Weizman says a Claude Dispatch design flaw lets a browser extension with access to claude.ai make Claude Desktop run commands; Anthropic closed his May report as a duplicate and has not fixed it. Docker Sandboxes 0.51.0, released on 3 November, fixes CVE-2026-103916, which let AI agents use symbolic links to reach files outside approved folders, including SSH keys. Mistral Vibe versions before 3.0.3 have a remote code execution flaw rated 8.6 that runs an attacker's program through git settings before a trust prompt appears. Claude Dispatch · Docker Sandboxes · Mistral Vibe

RunSybil says a CPython memory safety flaw lets malicious model templates escape Jinja sandboxes in AI inference servers; fixes went into Python 3.13, 3.14 and 3.15 but not 3.12. In AE Studio tests for the AI Alignment Foundation, coding agents posted a private token 15% of the time when a forum of 24 agents demanded it before helping. SecureFlag analyzed more than 800,000 attempts to fix software vulnerabilities and found that 25% were implemented incorrectly. CPython · AE Studio · SecureFlag

Claude Code's "don't ask again" option saved approved commands, secrets included, to a project settings file that leaked 2,539 developer secrets to public GitHub repositories, Apiiro found; Anthropic says it is looking at fixes. Claude Code secrets

AI models in security tests

In Vals AI's cyber range benchmark, GPT-6 Sol led with 27 of 31 objectives on the harder range, while open weight DeepSeek V4.1 Flash came close for about a tenth of the cost per run. Irregular CEO Dan Lahav said human error and a misconfigured test environment let models from Anthropic, OpenAI, Meta and Google reach the internet during evaluations, and that the models did not bypass safeguards. Google DeepMind says human guided AI agents found two weaknesses in legacy 1024-bit RSA; modern 2048-bit RSA systems are not affected. Vals AI · Irregular · RSA

CrowdStrike

CrowdStrike agreed to acquire agent observability company Fiddler AI and AI security testing start up Dreadnode, in deals reported at about $495 million combined; Fiddler's technology will be sold as Falcon Observe. CrowdStrike added an AI Kill Switch to Falcon Guardian that stops rogue AI agents, blocks them from relaunching and cuts their network access. Its Agentic Adversary Emulation, in early access, simulates attacks within customer approved boundaries and tests whether security fixes work. Fiddler AI · AI Kill Switch · Adversary emulation

CrowdStrike integrated Falcon with Amazon Bedrock AgentCore Guardrails and Amazon ECS Managed Instances to protect AI applications and agents against prompt injection and data leakage. CrowdStrike and Lovable brought Lovable workspaces into Falcon Shield, so security teams can find publicly accessible apps, sensitive data and access issues. AWS · Lovable

Deals and funding

Palo Alto Networks is in talks to acquire Israeli start up Clutch Security, which secures the credentials of AI agents and other non human identities, for about $500 million; the talks may not lead to a deal. Palo Alto Networks is also negotiating to acquire Beacon Security, an Israeli start up that builds AI agents for threat detection and response; terms were not disclosed. Doppel acquired Y Combinator backed PropRise, whose co founders join Doppel's engineering team. Clutch Security · Beacon Security · Doppel

Sentra raised a $150 million strategic investment led by Capital One Ventures for agentic data security, and a Capital One representative joins its board. Brava Security raised a $15 million seed round led by Pitango and F2 Venture Capital for AI agents that simulate attack logs to test whether security tools detect threats. Milan based CyberKord raised a €2.7 million pre seed round led by 360 Capital for its AI IT and cybersecurity platform. Sentra · Brava Security · CyberKord

Fourthline and Veridas are merging into one identity verification and fraud prevention company, and AshGrove is investing €40 million in Fourthline as part of a package of up to €65 million. Ballistic Ventures filed with the SEC to raise $360 million for its third flagship fund and $130 million for Ballistic Ventures Fund B. Fourthline and Veridas · Ballistic Ventures

Securing AI agents

Uber described its defence in depth framework for AI agents, with an agent registry, personas and intents, tool and MCP validation, runtime guardrails and continuous red teaming. Sysdig launched AI Defense, built on the Falco engine, which allows, confirms or blocks AI agent actions based on policies and keeps immutable audit records. Cyberhaven introduced Agentic Data Security, which discovers AI agents, records their prompts, tool calls and file reads, and applies data loss prevention policies. Uber · Sysdig · Cyberhaven

iboss announced AI DLP, which blocks sensitive data sent to AI tools and cloud apps without relying on data labels; a private preview starts in December 2026. Citrix expanded SecurSpaces to governed workspaces for employees and AI agents, with a free Community Edition and an open source Knowledge Worker Agent. Box introduced Box Architect for AI agent workflows and made agent security controls generally available to eligible customers. iboss · Citrix · Box

MPCH and Socura will test MPCH's ATLAS control plane, which stops AI agents from isolating devices, blocking users or changing settings without a human operator's approval. Kihan introduced Constitutional Computing, which checks cryptographic permits before AI agents take consequential actions; evaluations of its first product, Inktomi, are planned for December 2026. MPCH and Socura · Kihan

Identity for AI agents

LexisNexis Risk Solutions added Know Your Agent intelligence to its fraud prevention platform, so organizations can allow, review or reject AI agent interactions in existing risk workflows. Cloudflare launched analytics that identify personal agents visiting websites, plus access policies that let site owners control them without blocking agents for all users. Descope launched Envoy, which gives personal AI agents their own identities in consumer apps using the proposed Personal Agent Protocol; GoFundMe and GoodRx are partners. LexisNexis · Cloudflare · Descope

Zetrix AI and China's CAICT launched a Know Your Agent service that gives AI agents verifiable identities, delegated permissions and activity records, with a pilot planned across the Middle East, Africa, China and ASEAN. Zetrix AI

Cloudflare and Tools for Humanity proposed Trusted Agent Delegation, which ties AI agents to accountable users so websites can tell legitimate agent traffic from abuse. Trusted Agent Delegation

AI tools for security teams

Google Cloud added CodeMender features for isolated vulnerability testing, exploit simulation, human approval checks and targeted repository scanning, plus new agent security controls. PortSwigger made its agentic pentesting tool Burp AT available to enterprises, with controls for testing scope, permissions and automation levels. Assail introduced Cassandra, an AI agent in its Ares platform that monitors attack surface changes and recommends security tests, with human approval still required. CodeMender · Burp AT · Assail

Checkmarx is now on Anthropic's Claude Marketplace, so US customers can put part of their Anthropic spending commitment toward Checkmarx Fusion SAST, Triage Assist and Remediation Assist. Constructive released Safegres, an MIT licensed open source tool that audits PostgreSQL grants and row level security policies before AI generated applications are released. StrongestLayer launched Intent Workflows, which holds high risk emails involving payment fraud or AI prompt manipulation for human review. Checkmarx · Safegres · StrongestLayer

Atlantic Data Security partnered with LogSeam for its Harbour AI SOC managed service, which uses agentic AI for routine security operations under human oversight. Ontinue launched MXDR Essentials, which extends Microsoft's Integrated Security Operations Center with agentic AI and 24/7 human support. Atomicwork launched a marketplace of certified, prebuilt AI Coworkers for enterprise IT and security operations. Atlantic Data Security · Ontinue · Atomicwork

Conduktor added free AI agents to its Console platform that monitor Apache Kafka environments, identify risks and recommend fixes for human approval. Conduktor

Google Cloud updated CodeMender, which validates vulnerabilities and generates tested patches, and added identity controls and inspection tools for AI agents in production. Google Cloud

Governments and policy

Japan announced a 10 billion yen package to share cyber threat information, deploy response teams, use AI to find network weaknesses and support active cyber defense. An amendment to the UK Cyber Security and Resilience Bill would let ministers set rules on AI agent safeguards, incident reporting and investigations at essential service operators and cloud providers. The UK National Cyber Security Centre says AI will mainly speed up and scale existing attack methods, and organisations with weaker security will face greater risks. Japan · UK bill · NCSC

Malaysia will use communications and cyber safety laws to investigate an alleged foreign linked AI influence campaign targeting voters, and MCMC has requested technical evidence from Anthropic. Capgemini will lead a consortium with Wavestone and France's LNE to help the European Commission's AI Office assess cyber risks of advanced general purpose AI models. A CNAS report warns that sovereign AI agents that can change their goals, reproduce and resist shutdown could become impossible to control. Malaysia · Capgemini · CNAS

NIST published IR 8621, a report on its 14 May 2026 workshop on managing superintelligence incidents, including attacks, misuse and malfunctions. The University of Nebraska-Lincoln will lead cybersecurity for the five year, $24.5 million Fabric for AI-Driven Science project selected by the US National Science Foundation. Stream Engine delivered a deepfake detection analytic to the Air Force Research Laboratory's test and evaluation testbed under a research agreement. NIST · Nebraska · Stream Engine

Wraithwatch earned FedRAMP High certification through Knox Systems, so US federal agencies can deploy its AI cyber defense platform. Finch AI received CMMC Level 2 certification after an independent assessment by Ace of Cloud. Darkforce says its cyber and AI security platform was assessed as Awardable in the CDAO Tradewinds Solutions Marketplace for US Department of War work. Wraithwatch · Finch AI · Darkforce

Nvidia's board opened an independent investigation into chief executive Jensen Huang after reports that he was targeted by Chinese intelligence; a US intelligence memo claimed a Megaspeed executive was recruited to spy on him, reports say. Board investigation · Intelligence memo

Also

Dubai based bank Mashreq added generative AI tools to its fraud management platform with Clari5. Hong Kong based iCOMPASS introduced KY-AI, which helps financial institutions identify AI systems, assign accountability and monitor changes. Aethereal Cybersecurity and Global Cybersecurity Alliance are setting up a Centre for AI and Cybersecurity Research in Hong Kong. Mashreq · iCOMPASS · Hong Kong centre

SensAI Veritas launched a platform for risk screening and deception detection at border, airport and prison checkpoints; the company says its risk indicators are not proof of deception. SensAI Veritas

Weill Cornell says patient records from its imaging service were processed with AI tools without authorisation between 10 and 20 June, and it has suspended the research. Weill Cornell