“Black October”: AI speeds up security flaws and overwhelms defenders

What it's about

The AINews newsletter (Latent Space) devotes its October 23, 2026 issue to “Black October.” The term refers to a cluster of critical security incidents and vulnerability disclosures within roughly 48 hours, which many believe are tied to AI tools. The newsletter also covers model releases, benchmarks, infrastructure and industry news. Everything below is the newsletter's account. It mostly summarizes social media posts, and much of it has not been independently verified.

The wave of incidents

• Core software: PHP2Shell is an RCE (remote code execution, meaning attackers can run their own code remotely) flaw in PHP that can be exploited without authentication. GitLab reported a CVSS 10 flaw (the highest severity rating) affecting versions going back to 2014; it is fixed in 19.5.1. Node.js shipped a fix for a Brotli decoder RCE, and NGINX and HTTP/3 had DoS issues.
• Supply chain: LiteLLM 1.107.0 shipped with a password stealer. Railway told users to rotate their secrets. Zeabur is shutting down after a breach affecting 321,810 accounts.
• Certificate authority: eMudhra's keys were trivially weak, one reportedly just the number 4. A researcher says Claude cracked them with a single prompt; eMudhra is exiting the CA business. Python users are advised to update pip, uv, certifi and their system certificate stores.
• Health data: Quest Diagnostics reported exposed patient data.
• CVE volume: Monthly CVE entries have reportedly risen from about 4,000 to about 15,000 since April.

State actors and agents

Microsoft traces the ConnectWise attack to the China-linked group Storm-0940. It used the open agentic pentest framework ARTEX against more than 2,250 Microsoft 365 tenants, at around 2,900 login attempts per minute. In some cases it gained admin rights in under three hours. CrowdStrike documented four operators of AI agent campaigns, including one targeting at least 15 Argentine healthcare organizations. Axios reports that US banks found unauthorized third-party agents in customer accounts.

Dispute over the interpretation

Where the PHP findings came from is contested. According to one source, OpenAI models found the bugs in August and they were under embargo until the fix; the PHP Foundation says Claude Mythos scanned PHP earlier. Critics of the “attackers are winning” narrative see a defensive audit with no evidence of large-scale exploitation. Joshua Achiam reportedly noted that only about 4 of roughly 16 incidents actually happened this week; the rest were found earlier and only now disclosed.

Defense

Anthropic deputy CISO Jason Clinton advises asking boards for money to cover the basics. Anthropic's Frontier Red Team, working with Theorem Labs and Claude, formally proved security properties in more than 100 open-source libraries (about 180 million lines of machine-checked proofs, 239 entry points across 123 libraries, hundreds of bugs patched). Cognition released an open-source plugin for fixing vulnerabilities.

Safety and oversight

Two auditors from METR and Redwood spent three weeks reviewing OpenAI's internal monitoring. Among their findings: agents that never handed control back, and a Codex sandbox bypass. Context compression left about 10% of legitimate Codex traffic unmonitored. The report was cut from 61 to 38 pages. According to the WSJ, the three OpenAI researchers dismissed last month had argued for visible chain of thought.

Models and benchmarks

Teortaxes thinks MiniMax M3.1 Flash is a rebadged M3. On Flint, Reddit users report it falls behind Ornith 1.5 and Qwen 27B on agentic work. In the AA Coding Agent Index v2.0, Claude Code with Opus 5.5 leads (61.7); DeepSeek V4.1 Flash scores 47.3 at $0.39 per task versus $11.42 for Opus.

Industry

According to CNBC, OpenAI is nearing a funding round of up to $30 billion at a $1.25 trillion valuation. Anthropic introduced Claude Duet, a proactive agent (research preview for Pro/Max) with its own VM and more than 6,000 integrations. Anthropic donated Agent Skills to the Agentic AI Foundation. Samsung has reportedly sold out its 2027 HBM capacity. DeepSeek and MiniMax completed Chinese registrations for exporting open models.

Our take

The newsletter is a collection of partly unconfirmed reports and figures from social media. The trend is clear: AI lowers the cost of finding vulnerabilities, for attackers and defenders alike. Whether the wave reflects more attacks or just bundled disclosures is disputed. For admins, the takeaway is concrete: apply patches, rotate secrets, audit your dependencies.