AI work, server fixes and customer notices: October’s changes and evidence

The October 19 weekly covered choices including Meta's memory and sustained-work models. Hugging Face's October 20 AnyLanguageModel 1.0 was a developer library connecting Apple apps to models, rather than a model itself. Anthropic's October 22 Agent Skills donation concerned a format for passing task instructions. Calling all three a new AI obscures what an operator must prepare. Meta original · Library announcement · Skills announcement

OpenAI's GPT-6.1 Terra, introduced October 26, emphasized visual understanding and computer work. The model chooses an action; a connected tool performs the click or entry and returns the changed screen. A vendor's capability claims still leave the organization to define permitted actions. Terra announcement · Computer-use guide

In this fictional invoice task, steps 1–3 repeat until an entry draft is complete. A person then compares it with the original, approves saving and inspects what was actually saved.

%% Fictional screen-entry cycle · repeat 1–3 until the draft is ready; saving needs later human approval
flowchart TD
  A["1. AI reads document and screen"] --> B["2. Choose next field and value"]
  B --> C["3. Tool enters value<br/>Returns changed screen"]

PHP runs server-side processing for sites such as WordPress. Fixes for two remote-code-execution flaws were released October 21; US cybersecurity agency CISA added both to its Known Exploited Vulnerabilities catalog, or KEV, on October 26. A request reaching vulnerable processing can lead to attacker-chosen commands executing on the server. PHP announcements · CISA alert

The app observation belongs to the historical application follow-up, not a new connection for this monthly. The kernel is the operating system's core that manages programs and resources. An affected Linux kernel can retain a separate privilege-escalation flaw after PHP is fixed. If an attacker first gains limited execution inside the server and the additional conditions hold, access can expand to the highest administrator privilege. Ubuntu kernel guidance

%% Conditions for a chain · CISA KEV and Ubuntu, not an observation of a reader server
flowchart TD
  A["External request reaches<br/>Vulnerable PHP processing"] --> B["If flaw conditions hold<br/>Commands run as server account"]
  B --> C["Affected kernel and<br/>Additional attack conditions"]
  C --> D["Can reach administrator rights<br/>Server files and settings at risk"]

The Ivanti VPN story reported October 31 illustrates the same response split. Fixing the organization's access gateway does not investigate accounts and data already copied. CISA calls for network and memory investigation and warns that a disk checker can miss memory-executed activity. Ivanti response alert

GMO Internet's Onamae.com Desktop Cloud provides a Windows computer operated over the internet. After its compromise, the company said it began contacting subscribers about goodwill receipt and a replacement service at 15:30 JST on October 30. Company contact starting is confirmed; individual payment and migration completion are not. Customer notice

The October 24 policy separates ¥5000 goodwill from trading-loss compensation. The replacement service offers six free months after migration, with advance conditions and customer consent. Application methods, documents, period and payment timing for trading losses are promised separately by November 6. Notice 4

Internet infrastructure and security provider Cloudflare's security-audit-skill provides instructions and tools for a coding AI to find candidates, have another verifier challenge the evidence, and record confirmed, unresolved and disproved findings. We revisited the project from the October 26 article as additional reporting. Maintainer README

The figures have not been established to use a comparable measurement interval, so subtracting them cannot establish additional growth. Current monthly display · Commit · Releases

Code is MIT-licensed, with notice-preservation conditions. Execution requires a coding agent supporting tools and parallel agents, Node.js validators, and an OS-enforced sandbox that disables external networking and restricts writes and resources. Without those controls, the workflow retains the unresolved fact rather than running target code to confirm it. We did not install or execute an audit for this report. License

The weekly's fictional waiting-time example shows why the total matters. Data lookup takes 800ms, AI judgment 200ms and display 100ms; only judgment becomes ten times faster at 20ms. One ms is 1/1000 second. This is the weekly's explanatory calculation, not a measured product comparison.

Total reduction is 1.10 − 0.92 = 0.18 seconds, or 0.18 ÷ 1.10 ≈ 16.4%. Likewise, a monitoring feature alone does not establish that a transfer or deletion stopped. OpenAI's current API guidance separates alerts-only configurations from connections that can block continuation, and says stopping does not undo completed actions. Current monitoring guidance