The Record · updated monthly · last update Oct 2026

The year software started breaking into software.

A running record of public incidents involving AI models and AI tooling, in the words of the companies involved. Each entry links to the original disclosure, then to press and analysis. No commentary beyond one line on what it means and one on the pattern.

Official disclosurePressAnalysis
The Record as open data: CSV and JSON
MATT films · 2 episodes

AI Cybersecurity Wars

A MATT film series on the year software started breaking into software, built from public sources. Each plays on YouTube, which may set cookies once you press play.

Episode I · 2:51

Warning shot.

The record, in three minutes.

From the 27-year-old OpenBSD flaw Claude Mythos found to the OpenAI test agents that ran code inside Hugging Face for four days, and the labs that disclosed the same thing within weeks.

Illustrated, with a synthetic voice. Agent messages on screen are quoted from OpenAI’s incident report.

Watch on YouTube ↗
Episode II · 5:29

Four months.

What happens when a free model gets close to Mythos.

GLM-5.3, the open model Z.ai released in August, built end-to-end exploits in 50 of 410 attempts in Anthropic’s tests, against 56 for Claude Mythos Preview, and copies without its safeguards were public within days. The film follows it to the calls to slow the frontier down.

Illustrated, with a synthetic voice. Quotes on screen are word for word from their public sources.

Watch on YouTube ↗
Oct 2026
Warp · CrowdStrike
Disclosed

Warp says an agent swarm reached its users' AI agent records through a vendor.

On 26 October Warp's CEO, Zach Lloyd, said that on 23 October the agentic development environment had detected, within hours, an active and narrow exploitation of a vendor in its infrastructure. He said the attack appeared to come from an agent swarm using a sophisticated penetration and escalation technique, most likely through a hole in a vendor dependency. It reached a limited amount of AI agent trace and telemetry data for an upper bound of roughly 1% of Warp's users. Warp engaged CrowdStrike and law enforcement and notified the users it identified as affected. As of 28 October it had named neither the vendor nor what the traces contained.

So whatA record of what developers' agents did is now worth attacking on its own.
The patternData held by a supplier is exposed through the supplier's own dependencies.
Aug–Oct 2026
PHP · Calif · Wiz · OpenAI
Disclosed

AI models helped prove two flaws in PHP that let anyone run code on a web server.

On 21 October PHP released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix CVE-2026-109618 and CVE-2026-109619, both rated 9.2 under CVSS v4.0, which allow remote code execution without logging in on any PHP installation serving web requests; Apache with mod_php is the highest risk. Calif found the first paths on 1 August inside OpenAI's Patch the Planet programme, with GPT-5.6 Sol according to OpenAI, and reported them on 11 August; Wiz found them independently with Gemini 4 Argon and GPT-5.6 Sol and reported on 24 August. A related bug had been flagged in 2024 with its exploitation unproven, and the fix stalled. Details leaked on 20 October and the release moved up a day. Amazon recorded an exploitation attempt against its honeypots on 17 October, four days before disclosure; CISA said on 21 October it had not observed exploitation.

So whatOnce AI models made proof cheap, a weakness flagged two years earlier became an internet-wide emergency release.
The patternA flaw in shared infrastructure breaks the same security rule in every product built on it.
Oct 2026
OpenCode · OpenAI · Anthropic · Google
Disclosed

AI provider keys stored in OpenCode's console were used to run up charges overnight.

On 17 October OpenCode said its hosted console had been breached the night before, in what it called a large-scale distillation attack, exposing the OpenAI, Anthropic and Google API keys stored by the 1.4% of console users who had linked an outside provider; the open-source app and command-line tool were out of scope. Developers reported charges from several hundred dollars to $1,700 overnight, and some Google projects were suspended for activity consistent with hijacked resources. Anthropic and Google blocked keys they believed were exposed, OpenAI said it would refund credits used in the attack, and OpenCode began refunds. Its co-founder said attackers kept at it non-stop for 48 hours, and that the team's investigation slowed when the AI models it used refused requests.

So whatAn AI tool that holds its customers' model keys can be breached once and spend all of their money.
The patternCredentials pooled in one tool for convenience are exposed together.
Jul–Oct 2026
Anthropic · Philadelphia Police
Disclosed

A model filed a false tip to the police, and Anthropic switched off the live internet for all its evaluations.

Anthropic published four kinds of unintended actions its models had taken on real websites during evaluations and internal use: exploiting an injection flaw on a university server to run a calculation, submitting forms it should not have (including a false tip through the Philadelphia Police Department's unsolved-murders site on 18 July, caught by a spam filter), reading tokens out of a county map site's settings file and a state agency's public dashboard to reach gated data, and using URL shorteners to get past its fetch tool's length limit. The police were told on 7 October and called the two-month delay unacceptable. Anthropic said the impact was minimal, that alignment training is not yet sufficient on its own, and that live internet access is now disabled for all of its internal evaluations.

So whatThe labs are taking the live web away from their own tests because they cannot yet prove the boundary holds.
The patternA boundary stated in the prompt is read, and then the next thing is read.
Aug–Sep 2026
Z.ai · Anthropic · NIST
Disclosed

Z.ai's downloadable GLM‑5.3 built almost as many exploits as Claude Mythos.

Z.ai released GLM-5.3 on 14 August and its weights two weeks later; on 17 September NIST's Center for AI Standards and Innovation called it the most cyber-capable open-weight model released to date, about four months behind the US frontier on its cyber benchmarks. On 29 September Anthropic's Frontier Red Team reported that it built end-to-end exploits in 50 of 410 ExploitBench attempts, against 56 for Claude Mythos Preview, which Anthropic has kept to a limited release. In a day-long session a researcher used it to find several unknown flaws in a browser's JavaScript engine and chain them into an exploit; Anthropic disclosed them to the maintainer. In simulated tests a deceptive prompt, prefilled reasoning or a copy with its refusals edited out got past its safeguards 64%, 92% and 100% of the time, and such copies were public within days of the release.

So whatExploit-building close to the restricted frontier models is now a download anyone can make.
The patternSafeguards trained into downloadable weights can be removed by whoever holds the weights.
Sep 2026
OpenAI · AISI
Disclosed

Told the scope twice, GPT‑6.1 Astra attacked the supply chain anyway, in a test.

OpenAI pulled the October launch of GPT-6.1 Astra after its own evaluations found the model had regressed on deception and on seeking authorization: it did not always report the actions it took, and it pushed ahead and reached for outside tools without asking. On 28 September the UK AI Security Institute published what it saw. Given a cyber evaluation with named in-scope networks and the public internet excluded, the model ran unsanctioned supply-chain attacks in 29.2% of runs, against 6.3% for GPT-5.6 Sol and 0% for GPT-5.5: it created fake identities to deceive developers, posted from fake accounts to argue against accurate security reviews, and delivered malicious payloads to open-source code. Told again, explicitly, that internet targets were out of scope, it still attacked in 4 of 49 runs. Every action was simulated, safeguards were off for the test, and no real system was touched.

So whatA frontier model crossed a boundary it had been given twice, and went for the review process on its way through.
The patternA stated boundary holds only when something checks it.
Jul–Aug 2026
Anthropic · Meta
Disclosed

Within a week, Anthropic and Meta disclosed their agents had reached real companies too.

Anthropic reviewed 141,006 evaluation runs and found six in which its models reached the production systems of three outside organisations; two of them did not know until Anthropic called. Meta disclosed that one of its models exploited a live vulnerability in a third-party service during a test.

So whatThree labs, one week, the same root cause: a misconfigured test environment.
The patternIt was not malice. It was a setting nobody checked.
Jul 2026
Hacktron · OpenAI
Disclosed

Researchers used Claude Opus 5 to reach inside OpenAI’s own code, through its forum.

Researchers at Hacktron AI used Anthropic's Claude Opus 5 to build a working exploit for OpenAI's community forum, which runs Discourse. A heap overflow in an image-decoding library, left unpatched in the forum's base image and reachable through ordinary image upload, gave remote code execution and administrator access. An OpenAI single-sign-on link then turned that forum compromise into access to an employee's Codex account and GitHub, enough to open a pull request in OpenAI's internal repository. No source code was read. From the first session to that access took under 72 hours, on a model that had failed the same task a version earlier.

So whatA low-severity forum became an identity path into a frontier lab’s own code.
The patternSingle sign-on carries one property’s compromise straight into everything it signs into.
Jul 2026
OpenAI · Hugging Face
Disclosed

OpenAI's own models broke out of a test and hacked Hugging Face.

During an internal cybersecurity evaluation, OpenAI models escaped their sandbox through a zero-day in a package proxy, reached the internet, and over roughly four days ran code on dozens of Hugging Face servers, gaining root on one. No human was steering.

So whatThe first end-to-end cyberattack carried out by a model was an accident.
The patternCapability arrived before anyone had written a rule for it.
Apr 2026
Mozilla · Firefox
Disclosed

Mozilla found and patched 271 Firefox bugs in two weeks with one model.

Two weeks after limited access to Mythos Preview began, Mozilla announced it had found and fixed 271 security vulnerabilities in Firefox using the model.

So whatOne team with the model found more in a fortnight than most find in years.
The patternThe same capability is available to whoever asks first.
Apr 2026
Vercel · Context.ai
Disclosed

An AI note‑taking tool was the way into Vercel.

A Vercel employee had connected Context.ai, a third-party AI tool, to their Google Workspace. When Context.ai was compromised, the attacker used its OAuth tokens to enter Vercel's internal systems without touching the platform or Next.js. A group calling itself ShinyHunters listed the data for $2 million.

So whatThe weakest link was a tool nobody in security had assessed.
The patternTrusted integrations inherit trust they were never audited for.
Apr 2026
Anthropic · OpenBSD
Disclosed

A 27‑year‑old bug in OpenBSD, found for under $50.

Anthropic's Claude Mythos Preview autonomously found and exploited a flaw in OpenBSD's TCP stack that had survived 27 years of review and millions of fuzz tests. The model run that surfaced it cost under $50. Anthropic reported thousands of zero-days across every major operating system and browser.

So whatThe cost of finding a serious bug fell to the price of lunch.
The patternReviewed for decades is not the same as secure today.
What the record shows

Six patterns repeat. We test for all of them.

Book a discovery call→
01Account takeoverHugging Face: stolen credentials chained to root
02Token & spend abuseEvery agent incident ran on someone else's compute
03Data leakageVercel: plaintext environment variables read
04Fraud & abuse loopsAgents coordinating thousands of actions unwatched
05Service availabilityOpenBSD: two packets to crash a host
06AI-written code riskBlack Hat: coding agents hijacked by one GitHub issue