AI and software flaws
OpenAI says GPT-6 Cyber found 81 new high-severity flaws in Firefox. A PostgreSQL developer says bugs found with AI are delaying PostgreSQL 19. Google says some cyber defenders will get Gemini 4 Carbon first.
AI used to find or exploit weak spots in ordinary software and systems, AI used by defenders, and the patching work that follows.
November 2026
9–15 November 2026
7 stories in the AI newsletters I read.
The facts
Google DeepMind said its AlphaProof Code verified key security properties of Signal's cryptographic protocol.1
Arctic Wolf Labs reported that attackers used AntiFox, an AI hacking tool built for security tests, to steal AWS credentials.2
2–8 November 2026
10 stories in the AI newsletters I read.
The facts
OpenAI says GPT-6 Cyber found 81 new high-severity flaws in Firefox.1
A PostgreSQL developer says bugs found with AI are delaying PostgreSQL 19.2
The stories
- OpenAI launches Daybreak Defense Factory and Daybreak Red Team.
- Arctic Wolf Labs reports attackers use AntiFox, an AI pentest platform forked from ARTEX, to steal AWS credentials and query production services in victim environments.
- Daylight raises an $80 million Series B led by Goldman Sachs Alternatives for AI security operations.
- Devin adds Mythos 5.2 to security scans (73.5 vs 56.9 for Mythos 5.1) plus $100M+ in credits.
- Google DeepMind introduces AlphaProof Code, a multi-agent framework that mathematically verified key security properties of Signal's cryptographic protocol at the implementation level.
- OpenAI brings in security firm Hacktron to hunt vulnerabilities in its own systems and integrations.
- Cognition chief security officer Alex Stamos tells The Verge's Decoder podcast that cybersecurity has entered a dark age as AI helps attackers find vulnerabilities faster than defenders can patch.
- Vals CyberOffense finds GPT-6 Sol, available only via OpenAI's restricted Daybreak program, completes 80.8% of objectives; open-weight DeepSeek V4.1 Flash completes 72.8%.
- DeepMind agents with human experts find two weaknesses in legacy 1024-bit RSA; RSA-2048 is unaffected.
- India's cybercrime agency warns financial firms about agent-driven attacks, following a central bank alert (Reuters).
- OpenAI introduces GPT-6 Cyber, rated Critical for cyber, which built exploits for 948 Linux kernel bugs, found 81 new high-severity Firefox flaws and stayed in scope 99.93% of the time.
- OpenAI says GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities in Firefox; The Register reports Mozilla has started shipping fixes.
- OpenAI says a pipeline run with an upcoming unreleased cyber model validated and patched 948 Linux kernel local privilege escalation bugs at a total API cost of $1.2M.
- OpenAI says GPT-6 Cyber stayed inside its authorization boundary over 99% of the time, with 0.07% of cases out of scope, and does not say what happened in those.
- The UK's NCSC warns an AI-powered ransomware attack on British critical national infrastructure is 'highly likely' by 2028.
- OpenAI reprices its bug bounty: midpoints rise more than 4x on average, critical findings pay a base up to $30,000, and bonuses can lift one award to about $172,000.
- A PostgreSQL contributor says PostgreSQL 19 will likely slip as back-patches jump over 60%.
October 2026
26 October – 1 November 2026
14 stories in the AI newsletters I read.
The facts
Google says some cyber defenders will get Gemini 4 Carbon first.1
The facts
Bank of America's CEO says the bank added 1,000 contractors to patch flaws exposed by AI models like Anthropic's Mythos. It already had 3,000 staff doing this work.1
OpenAI says its models helped find critical flaws in PHP, and that it spent two months preparing fixes. A note added by X readers says an independent researcher found the flaw first.2
Anthropic says the first scans of its free OSS Scanner sent open-source maintainers 77 critical and 803 high-severity findings. Its models set these ratings.3
The stories
- OpenAI leads a coalition patching critical PHP vulnerabilities in a coordinated disclosure.
- Dream publishes research, verified by Palo Alto Unit 42, on attackers chaining PHP2Shell with Fragnesia to root 25 servers in under 15 minutes per chain.
- CISA adds the Fragnesia Linux kernel flaw (CVE-2026-46300) to its Known Exploited Vulnerabilities catalog.
- CrowdStrike attributes a campaign to China-nexus VAPOR PANDA using an agentic framework called ARTEX, targeting more than 200 organizations in at least 14 countries within 11 days.
- Epoch finds GLM 5.3 autonomously built Firefox exploits for six recent versions at about $1,000 per version; Chrome attacks did not reproduce.
- A supply chain attack involving ARTEX, an open source agentic AI penetration-testing framework, helped attackers reach more than 2,250 Microsoft 365 tenants through a ConnectWise compromise.
- Researchers document attackers using AI agents to obtain password hashes and MFA secrets from a European government system within two hours.
- Researchers identify AI-using attack infrastructure that deployed 44 agents to probe at least 19 Japanese organisations.
- Google DeepMind announces Gemini 4 Carbon, claiming 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench; access goes first to trusted cyber defenders.
- Anthropic says an unreleased Claude research model chained flaws into working exploits across Chrome, Safari, Android, Windows and Linux, including a 9.8-rated zero-click Linux chain; all bugs were patched before disclosure.
- Cathedral, a startup from ex-DOGE staffers, says its AI agents reached parts of Anthropic's internal systems within 12 hours.
- Cognition launches Devin Security on Claude Mythos with "hundreds of millions" in credits; CodeRabbit joins OpenAI's Daybreak network.
- OpenAI CEO Sam Altman says at Oracle AI World that he is shocked at how many vulnerabilities AI systems are finding and that the only way through is to find and fix them.
- Oracle announces the OpenAI-powered Oracle AI Security Mesh, plus OCI Identity Assurance, OCI Zero Trust Packet Routing and a wider collaboration linking ChatGPT and Codex to Oracle.
- OpenAI discloses an unauthenticated remote-code-execution vulnerability in Brotli, fixed in source in February but not flagged as security-relevant; no release carries the fix.
- OpenAI describes about two months of coordinated remediation before disclosure of PHP2Shell, calling it a Defender's Window.
- Cloudflare deploys WAF mitigations for PHP2Shell before the patch, free tier included; its CTO calls them a stopgap.
- The NSA advises preparing for an immediate PHP update and taking systems offline if mitigation is not possible.
- OpenAI publishes a Codex skill repository to help operators check PHP2Shell exposure.
- SemiAnalysis reports 18 Japanese companies disclosing breaches since September 25, none tied to PHP2Shell.
- eMudhra exits the CA business after its CA keys are found trivially weak; a researcher says Claude cracked them with a single prompt.
- CrowdStrike documents four operators running AI-agent campaigns, including one hitting at least 15 Argentine healthcare organizations.
- Anthropic's Frontier Red Team, with Theorem Labs, uses Claude to prove security properties across more than 100 open-source libraries, with about 180M lines of machine-checked proofs.
- DeepSeek V4.1 Flash rediscovers 87.5% of benchmark CVEs at pass@3 versus Astra at 90.6%, at $2.40 per CVE versus $135.23 for Astra.
- OpenCode reports a 48-hour attack from agent swarms; frontier models reportedly refused to help analyze the payloads.
- Anthropic's OSS Scanner offers free periodic vulnerability scans of critical open-source projects; first week yields 77 critical and 803 high model-rated findings, including a curl fix.
- Red Hat CTO Chris Wright says cybersecurity-focused AI models since mid-2026, starting with Mythos and Glasswing, have uncovered vulnerabilities latent in code for years, some for two decades or more.
- Bank of America hires 1,000 contractors to patch holes that Anthropic's Mythos exposed, CNBC reports.
- Reuters reports NorForge researcher Snorre Fagerland watched a hacker use AI agents on Cairn to break into websites of more than 100 companies, at about $50 per successful hack.
- Claude traces the NIST ECDSA curve seeds to a timestamp-seeded RNG.
- A model in RL training reportedly exploits unchecked token signatures on a government site to pull server secrets and data.
- Epoch estimates open-weight cyber capability trails the closed frontier by 6-7 months; GLM-5.3 scores 160 on its Cyber ECI, above GPT-5.4.
- OpenAI reportedly moves 25% of its production engineers to security work.
- Cohere acquires Aisle, a defensive-cyber agent company credited with 200+ high/critical CVEs; Cohere plans air-gapped deployments.
- XBOW reports a Google-confirmed $250,000 full-chain Chrome bounty; humans and agents using GPT-5.6 Sol built the exploit in six days with $40,000 of compute.
- Linus Torvalds says he "did what I preach", using LLMs to untangle kernel code and catch three bugs from 2013.
- Gemini 4 Argon is reported at 77.9% on DeepSWE v1.1 vs Opus 5.5's 74.2%, shipping first to 650+ Fairwind Program defenders at $2/$10 per M tokens.
- Anthropic's new Cyber Mission includes OSS Scanner, offering free periodic vulnerability scans of opted-in open-source projects.
- A CrowdStrike report attributes the attack on South Korean banks possibly to a single person, using a stack including DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6 and Claude Code.
- Anthropic expands its Cyber Verification Program to Mythos 5.1, Opus 5.5 and Sonnet 5.5, adding tiers for authorized offensive work.
- Google launches Gemini 4 Argon with 1M output tokens, first for cyber defenders.
September 2026
Gemini 3.8 Flash Cyber went only to trusted defenders. Anthropic says the GLM-5.3 model can build exploits on its own.
6 stories in the AI newsletters I read.
The facts
Gemini 3.8 Flash Cyber went only to trusted defenders.1
Anthropic says GLM-5.3 can build exploits on its own.2
All 6 stories
- Anthropic finds Zhipu's open-weight GLM-5.3 builds exploits nearly as well as Claude Mythos Preview, with safeguards removable for $4,400.
- Three researchers use Claude Opus 5 to chain bugs into a ChatGPT and Codex account takeover in under 72 hours.
- OpenAI describes Defense Factory, a 250+ person effort using models to find and fix vulnerabilities across hundreds of systems.
- OpenAI commits a $1B Daybreak subsidy and access program for defenders and critical infrastructure.
- Google ships Gemini 3.8 Flash Cyber for defenders only, reporting 86.2% on CyberGym and 47.2% on CWE-Bench patching.
- Threat actors build AI agent attack frameworks that need no human in the loop, so far using known flaws rather than zero-days.
August 2026
OpenAI called for a collective cyber defense. Its staff told Black Hat that their models' attacks are an urgent warning.
11 stories in the AI newsletters I read.
The facts
OpenAI's call for collective cyber defense urges giving more defenders cyber-capable AI.1
OpenAI staff told Black Hat their models' attacks are an urgent warning to the AI industry.2
All 11 stories
- Bank of England governor Andrew Bailey warns the G20 that autonomous frontier models could reprice cyber risk.
- AI-driven bug hunting pushes Linux kernel CVEs from about 500 per release toward 2,000.
- OpenAI publishes a call for collective cyber defense; more than 100 organizations, including Anthropic, Google and Microsoft, sign it.
- OpenAI's Greg Brockman urges security teams to use AI agents on old flaws before a near-frontier open-weight cyber model drops.
- Z.ai says its GLM-5.3 scores 84.5% on vulnerability discovery against Mythos 5's 83.8%, though it trails badly at building exploits.
- Reports say AI is compressing the time to find and exploit software weaknesses to minutes.
- Linus Torvalds releases Linux 7.2-rc7 with many fixes from AI tool review.
- OpenAI releases GPT-5.6 Cyber to a short list of partners including Accenture, IBM, CrowdStrike and Cisco.
- Meta cites HuggingFace handling a recent security incident by using widely available open models to patch vulnerabilities.
- OpenAI ships Codex Security Review in research preview for pull-request security checks.
- Apple caps vulnerability reports after AI submissions mixing real flaws with slop overwhelm its reviewers.
July 2026
Google's Chrome team says its June releases fixed 1,072 security bugs.
12 stories in the AI newsletters I read.
The facts
Google's Chrome team says its June releases fixed 1,072 security bugs.1
All 12 stories
- Chrome's June releases fix 1,072 security bugs, more than the prior 23 releases combined.
- OpenAI ships Codex Security, an Apache-licensed repo scanner.
- Linux kernel maintainers face 432 LLM-found CVEs in one weekend and expect a very long 18 months.
- Google limits Gemini 3.5 Flash Cyber to governments and trusted partners and says it found 55 confirmed V8 flaws, against 47 for 3.5 Flash.
- Hugging Face defenders turn to China's open-weight GLM 5.2 after US frontier models block their incident response.
- OpenAI reports GPT-5.6 Sol sets a cybersecurity record that already helps teams find and patch real vulnerabilities.
- The White House launches Gold Eagle, a frontier-AI clearinghouse for patching critical infrastructure vulnerabilities.
- Tracebit uses prompt injection as a defense that derails autonomous attackers.
- In JadePuffer, an AI agent runs a whole ransomware attack after a human picks the target; recovery looks impossible even after payment.
- The US cyber defense agency uses Anthropic's Mythos to scan federal code for vulnerabilities.
- Anthropic reports Claude Mythos Preview helped researchers find weaknesses in HAWK, a quantum-resistant scheme, and in AES-related algorithms.
- A threat actor abuses Gemini CLI to operate a botnet.
June 2026
Anthropic opened Project Glasswing to about 150 more organizations. Epoch AI estimates Mythos Preview was about 7 months ahead of trend on exploits.
10 stories in the AI newsletters I read.
The facts
Anthropic opened Project Glasswing to about 150 more organizations.1
Epoch AI estimates Mythos Preview was about 7 months ahead of trend on exploits.2
All 10 stories
- Chinese AI systems reportedly match Anthropic's Mythos in cybersecurity.
- OpenAI expands Daybreak with a Codex Security plugin and GPT-5.5-Cyber, which scores 85.6% on CyberGym.
- AWS ships Continuum, which it says fixes code flaws at machine speed, and Context, a knowledge graph for company data.
- Chainguard launches the Athena coalition, patching 2,000 open-source flaws before disclosure.
- Epoch AI judges the Mythos family a real leap in exploit development, seven months ahead of trend.
- Epoch AI starts tracking CVEs from every reporting org; high-severity disclosures spiked as Claude Mythos Preview landed.
- Anthropic expands Project Glasswing, opening Claude Mythos Preview to about 150 organizations.
- Anthropic gives the EU cybersecurity agency ENISA access to Mythos through Project Glasswing.
- Anthropic releases the Defending Code Reference Harness for finding and fixing vulnerabilities with AI.
- Researchers build an agent-driven worm that runs open-weight models on infected machines to craft attacks per target.
May 2026
The stories
- Greg Kroah-Hartman says AI bug-finders push Linux CVE issuance to about 13 a day.
- IBM and Red Hat commit $5 billion and announce Project Lightwell, a security clearinghouse for open source.
- Anthropic ships a security-guidance plugin for Claude Code, cutting security-related PR comments 30-40% internally.
- The ECB summons bank leaders over the systemic threat from Anthropic's Claude Mythos Preview.
- Anthropic says Project Glasswing and partners found over ten thousand high or critical vulnerabilities in a month using Claude Mythos Preview.
- Bugcrowd sees submissions quadruple in three weeks as bug bounties flood with AI reports.
- Linus Torvalds calls the Linux security mailing list almost unmanageable under duplicate AI reports.
- Claude Mythos Preview helps researchers find the first known Apple M5 memory exploit.
- A blog post says GPT-5.5 Pro one-shots Insane heap-pwn challenges, breaking the open CTF format.
- ExploitBench ranks Claude Mythos Preview first at 69% in reaching arbitrary code execution.
- Mythos finds five potential curl vulnerabilities, one of them legitimate, in code audited by humans and AI.
- Cloudflare publishes a report on Project Glasswing and Claude Mythos, noting Mythos chains vulnerabilities.
- Copy.fail, a Linux kernel privilege escalation bug, is exploited within a day of release and defeats container isolation.
- Linus Torvalds says AI tooling does much patch work as Linux submissions jump 20%.
- Microsoft ships a larger Patch Tuesday as AI scanning surfaces more bugs in its code.
April 2026
The stories
- OpenAI releases GPT-5.5 Cyber, a security-trained model limited to a small group of trusted users.
- The Firefox team uses Claude Mythos to find 271 previously unknown vulnerabilities in Firefox.
- The AI Security Institute evaluates Claude Mythos Preview and finds a step up in cyber capability over earlier frontier models.
- Google adds three agents to Google Security Operations: Threat Hunting, Detection Engineering and Third Party Context.
- NIST stops assigning severity scores to lower-priority vulnerabilities but still lists them in the NVD.
- AISLE finds small open-weight models can match Mythos at finding vulnerabilities.
- Claude finds zero-day remote code execution bugs in Vim and Emacs that trigger on file open.
- Anthropic withholds Mythos after it finds thousands of unknown vulnerabilities and launches Project Glasswing for restricted defensive access.
March 2026
February 2026
January 2026
November 2025
The board, once a week, by email.
It comes every Monday.
By subscribing, you agree to get the board by email.See the Privacy Notice.