Black October is the name developers gave to a run of breaches, backdoored releases and critical patches that landed almost on top of each other in October 2026. One weekly summary, published October 26 by the account MTS (@MTSlive, written by @theojaffee), packed the Oct 19-25 week into a single timeline and argued that AI cyber capability, which had climbed quietly for months, finally showed up in the real world.
We checked that timeline against what Beam has already published. Most dated events are confirmed and have their own breakdown. A few claims, mostly model benchmarks and executive quotes, we have not seen independently, and we mark them that way.
TL;DR
| Question | Short answer |
|---|---|
| What happened in Oct 19-25? | Cairn agents hit 100+ sites, PHP, GitLab and Node.js shipped critical fixes, Railway, LiteLLM and Quest Diagnostics disclosed breaches, GMO Internet was attacked. |
| Is it all AI? | No. Cairn and BUGSTER are agent-run. GitLab, Node.js and PHP are ordinary critical bugs, some found with AI help. |
| Is it verified? | Dated incidents: yes, see the table. Benchmarks and quotes from the thread: not independently confirmed. |
| Biggest action? | Patch, rotate secrets stored on shared platforms, and stop giving agents long-lived keys. |
| Does Beam stop this? | No. Beam watches what local coding agents do and flags risky actions. It does not patch your servers. |
Why did AI cyber capability take months to show up?
The summary's core claim is a timing one. Frontier models got better at offensive tasks through the spring and summer (it cites Anthropic's Mythos Preview training, then Project Glasswing, OpenAI's Daybreak, and Chinese open-weight models trailing by a few months), yet it reports no visible rise in attacks on US citizens between June and September. Then the incidents bunched up in one week.
A plausible reading, and it is only a reading, is that capability and exploitation have different lag times. Finding a bug is fast. Packaging an exploit, standing up infrastructure, and getting victims is slower, and it lands when the tooling is cheap enough to run at scale. Cairn is the clearest example: Reuters reported agents running through an open-source platform for a cost estimated near $5,000. Our Cairn breakdown separates what Reuters verified from what is an estimate.
What happened, day by day?
The days below follow the summary, with the Beam post for each incident. Dates are as the summary gives them unless the linked post says otherwise.
Monday, Oct 19-20
- Cairn: AI agents breaching 100+ sites, per Reuters and NorForge. See the Cairn claimed-vs-verified post.
- Anthropic announced a free scanning service for open-source repositories. Our take on continuous scanning and where it fits is in Claude Security continuous scanning.
- PHP2Shell, a critical PHP flaw pair, was disclosed. See PHP2Shell and the speed of AI-assisted exploitation and the patch guide.
Tuesday-Wednesday, Oct 20-21
- GitLab CVE-2026-110264, a CVSS 10.0 unauthenticated RCE for self-managed instances: patch details.
- Node.js security release, including the Brotli bug later traced to a shared library: Node.js October release and Brotli fixes beyond Node.
- Railway disclosed access to roughly 100,000 customer environment variables: rotate every secret.
- LiteLLM 1.107.0 was compromised on PyPI and Docker: LiteLLM compromise.
- OpenAI's disclosure program for AI-found bugs is analyzed in Patch the Planet.
Thursday, Oct 22
- Banks reported a jump in unauthorized agent activity on customer accounts, mostly users' own agents: US banks block agent logins.
- Quest Diagnostics disclosed unauthorized access to its Quanum LSM ordering system. It is in the October attacks timeline with a link to the SEC filing; the scope is still under investigation.
Friday-Saturday, Oct 23-25
- The Brotli origin of the Node.js bug, which also affects Bun, was published. The affected-software guide above lists fixed releases.
- GMO Internet, Japan's largest internet infrastructure company, reported an attack in which user financial accounts were drained: GMO breach, claimed vs verified.
For the full month, including everything outside this week, use the October 2026 attacks timeline and the supply chain incidents tracker.
What is verified and what is not?
| Claim from the weekly summary | Status | Where to check |
|---|---|---|
| Cairn agents breached 100+ sites | Reported by Reuters and NorForge; cost is an estimate | Cairn post |
| Critical PHP flaws patched | Confirmed, with patch guidance | PHP2Shell patch guide |
| GitLab RCE patch | Confirmed by GitLab | GitLab post |
| Node.js RCE patch from Brotli | Confirmed | Brotli guide |
| Railway breach, ~100k variables | Disclosed by Railway | Railway post |
| LiteLLM breach | Confirmed compromise of 1.107.0 | LiteLLM post |
| Quest Diagnostics breach | Disclosed in an SEC filing; scope under investigation | Timeline |
| GMO Internet attack | Confirmed by the company; ransom figures disputed | GMO post |
| DeepSeek-V4.1-Flash-1017 beats GPT-6 Sol on a Vals AI cyber eval | Not independently checked | Vals AI results |
| Kimi K3.1 Pro tops public models on cyber tasks | Not independently checked | Moonshot AI release notes |
| Abliteration sells a guardrail-free DeepSeek variant | Not independently checked; the idea matches the open-weight risk covered in the backdoored Qwen post | The vendor |
| Executive quotes (Altman, Prince, Clinton, Kumagai) | Quoted second-hand from the thread; GMO CEO's remark appears in our GMO post | Original interviews |
| Fortinet security updates "for agentic attacks" | Not independently checked | Fortinet advisories |
We have not run the models or inspected the vendor claims ourselves. If a vendor or researcher corrects one of these, we will update this post.
What do the numbers say about panic?
The summary itself offers the counterweight: SPY fell 0.95 percent over the week, a cybersecurity ETF rose 5.91 percent, and the VIX rose 5.82 percent, small next to past shocks. Most disclosures were fixes for bugs that were already patched, and defenders have far more resources than attackers.
What does change is cadence. Cloudflare's CEO is quoted saying there are more severe vulnerabilities in critical software that have not been disclosed yet. If that holds, the practical problem is not one big outage. It is a weekly stream of patches that small teams must apply faster than before.
How Beam handles this
Beam is a local, privacy-first monitor for AI coding agents such as Claude Code and Codex. It does not scan your servers or patch PHP. What it does for the part of this week that touches developers: it flags risky agent actions such as credential exposure and destructive commands, scans skills and MCP configs before you run them, and keeps its data on your machine. Enforcement is hook-based for Claude Code and Codex, not a kernel sandbox, and scanning is heuristic, not semantic malware analysis. If a stolen key is the way in, as with Railway and LiteLLM, the lesson is to limit what an agent can read in the first place. See the first 24 hours after a vendor breach runbook.
What should you do this week?
- Patch PHP, Node.js (and anything bundling Brotli, such as Bun) and self-managed GitLab. For GitLab, the October 12 patches do not fix CVE-2026-110264.
- Rotate secrets stored on Railway and any key that passed through LiteLLM 1.107.0. Do it before reading further logs.
- Turn on automatic updates for your OS, browser and apps. Use an authenticator app or a hardware key rather than SMS codes.
- Keep agent credentials short-lived and scoped. Do not leave long-lived API keys in shell environment variables an agent can print.
- Put a monitor on your running app so you see probing early, and keep an incident response playbook ready.
Related reading
- October 2026 attacks timeline
- AI-led attacks in October 2026
- Black October: every supply chain incident
- NIST, NSA and CISA agent control practices
- API keys in logs and env vars
Primary sources: the MTS weekly "Situation of the Week" post on X (Oct 26, 2026), and the linked Beam breakdowns, each of which cites its own vendor advisory or report.
Accurate as of October 27, 2026. This is a second-hand summary check, not a first-party investigation, and some items are still developing.