The Black October Security Prompt: What Claude and Codex Can Check
By Matt Cullerton, founder of Mavric and Apso
A TikTok from @nate.b.jones offers a prompt you can paste into Claude, ChatGPT or Codex to scan your machine and online presence for vulnerabilities tied to the "Black October" security event. Black October is real, and the advice to check your own exposure is reasonable. The prompt as spoken has four problems: it calls patched bugs zero-days, it asks models about an event after their knowledge cutoff, it assumes access these tools do not have by default, and it points an agent with file access at untrusted web content.
What is Black October?
Wikipedia records that October 2026 in information security "has been referred to as Black October due to several major security incidents and vulnerability disclosures." The incidents cluster between 17 and 21 October 2026:
- On 17 October, a researcher reported that the private keys for eight intermediate certificate authority certificates run by eMudhra were cryptographically weak.
- On 21 October, the PHP team released emergency fixes for CVE-2026-109618 and CVE-2026-109619, together called PHP2Shell, in PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12. WordPress, Drupal and Joomla sites were exposed downstream.
- A malicious LiteLLM version 1.107.0 was on PyPI between 11:39 and 13:23 UTC on 21 October.
- GitLab released a remote code execution fix on 21 October.
The video's opening claim, that AI is driving a wave of attacks, also has support. Fortune's 21 October 2026 report describes a 25 September attack in which one operator used autonomous AI agents against more than 1,200 websites and breached roughly 100 companies, including Deere, SharkNinja and Cummins, for about $5,000 in tokens.
Are the Black October bugs zero-days?
No. NIST defines a zero-day attack as "An attack that exploits a previously unknown hardware, firmware, or software vulnerability." PHP2Shell and the GitLab flaw were disclosed with fixes and CVE numbers on 21 October, five days before the video was posted.
The distinction changes what the prompt can do. An agent told to hunt zero-days on a laptop has nothing concrete to match against. An agent told to check whether the machine runs PHP below 8.2.35, 8.3.36, 8.4.27 or 8.5.12 has a version comparison it can perform.
Does Claude know about Black October?
Not without a web search. Anthropic's models overview lists Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 with a reliable knowledge cutoff of June 2026. Black October began on 17 October 2026, about four months later.
Asked about the event with no retrieval, a model can either say it does not know or produce something that sounds right and is not grounded in the real CVEs. In a security report, a confident list of invented vulnerabilities is worse than no list. The prompt never tells the tool to search first and never names the CVEs.
Can Claude or Codex scan your whole system?
Not out of the box. The video says "It will look at your whole internet presence. It really will. It will look at your whole system." Each tool limits that by default:
- Codex defaults to a
workspace-writesandbox withon-requestapprovals. OpenAI's documentation says "By default, the agent runs with network access turned off." It reads the workspace and writes only in the current directory. - Claude desktop reaches local files through extensions or folders the user connects. Cowork's file access is "limited to folders you connect."
- Claude Code asks before its file tools read or write outside the folder where it started, and does not auto-approve
curlorwget.
Getting the behaviour the video describes means deliberately widening permissions. The video does not mention that step.
Is the prompt itself a security risk?
It can be. The prompt tells an agent with filesystem access to read arbitrary web content about you, which is the setup both vendors warn about. Anthropic's 24 November 2025 research found prompt injection attack success rates of 7.6% on Claude Opus 4.5, 10.8% on Sonnet 4.5 and 11.1% on Haiku 4.5 before new safeguards, and 1.4%, 1.9% and 2.4% after. Anthropic wrote: "A 1% attack success rate, while a significant improvement, still represents meaningful risk."
Claude Code's security page advises "Avoid piping untrusted content directly to Claude." OpenAI's Codex docs say "Prompt injection can cause the agent to fetch and follow untrusted instructions."
The Black October incidents show what is at stake. Per Wikipedia, the LiteLLM payload "targeted browser credentials, cloud service secrets, developer authentication tokens and API keys used by AI services." A scanning session with broad file access and network access holds that same inventory in one place.
There is also a scope mismatch. Every Black October incident sits in server or developer infrastructure: PHP, Node.js, GitLab, a PyPI package and a certificate authority. None is a consumer desktop flaw. If you do not run a PHP server, self-hosted GitLab or a LiteLLM proxy, there is little Black October exposure for an agent to find.
How to make the prompt useful
Add one instruction before anything else: "search the web for the Black October CVEs disclosed between 17 and 22 October 2026 before you start, and cite each advisory you use." Then ask for specific checks, such as installed PHP and LiteLLM versions, instead of "zero-days." Keep the agent in a sandbox and approve any widening of access yourself.
Did people find scary results?
The video says people who ran the prompt came back "horrified." The video never shows the prompt running or any output, and no published account of a run turned up. Treat that claim as unverified.
Bottom line
Black October was a real cluster of patched server-side flaws, not zero-days. Current Claude models do not know about it unless they search the web, and none of the named tools scans your whole system by default. If you run the prompt, add a web-search step, name the CVEs, and keep the agent sandboxed.
Sources
- Black October (information security), Wikipedia - incident dates, CVEs, fixed PHP versions and the LiteLLM payload
- Fortune, 21 October 2026 - the autonomous-agent attack campaign
- NIST zero day attack definition - the definition used above
- Anthropic models overview - current Claude models and knowledge cutoffs
- Claude Code security docs - working directory boundary and untrusted-content advice
- Codex agent approvals and security - default sandbox and network-off setting
- Anthropic prompt injection research - measured attack success rates
- Installing Claude Desktop - how the desktop app reaches local files
- Nate B. Jones bio, CXOTalk - the creator's background in AI product strategy
Published October 26, 2026. Based on a TikTok by @nate.b.jones. Full video notes (markdown)