Skip to content
Daily BriefNewsDaily Brief

5 AI stories from October 22-24, 2026: Anthropic hands off Agent Skills, GitHub tightens Copilot governance, a harness-cost study, and two active security incidents

· by Pondero Newsdesk · 5 stories

AI news daily brief: 2026-10-24

Five stories cleared today's bar: one standards donation, one batch of platform-governance updates, one cost-benchmark study, and two active security incidents.

Anthropic donates its Agent Skills standard to the Agentic AI Foundation

Anthropic handed off Agent Skills, the SKILL.md format it built for packaging reusable agent instructions and scripts, to the Agentic AI Foundation (AAIF) on October 22, 2026, per Anthropic's announcement. The AAIF is the Linux Foundation body that already hosts the Model Context Protocol, Block's goose framework, and OpenAI's AGENTS.md. Agent Skills joined the same day as Hugging Face's OpenEnv, Alibaba's OpenSandbox, and OpenAI's Open Responses, bringing the foundation to 10 hosted projects in its first year. Anthropic says more than 70 platforms already support the format, including Microsoft Copilot, Cursor, GitHub, and Figma, with partner-built skills from Canva, Stripe, Notion, and Zapier live at launch. Anthropic's proposed roadmap for the standard, security guidance among it, has not shipped yet. That gap matters: a JFrog report published the day before the donation counted nearly 1,000 malicious skills already carrying harmful payloads. Full story: Anthropic hands Agent Skills to the Agentic AI Foundation as a report flags 969 malicious skills

GitHub ships five Copilot governance and usage-visibility changes in 48 hours

GitHub shipped a run of Copilot changes on October 22 and 23, 2026. "Global default for Copilot features" reached general availability, letting an admin set one default, Enabled, Disabled, or Let organizations decide, for every unconfigured eligible feature and client capability, including MCP server access, enforced in stages through October 26, per GitHub's changelog. The same day, the usage metrics API started reporting the configured names of custom skills, agents, MCP servers, slash commands, and plugins instead of grouping them under "other," per GitHub. On October 23, that API gained a new breakdown covering tools that skip editor telemetry entirely, among them OpenCode, Claude Code, Codex, and Zed. GitHub also added an opt-in industry-comparison view to the Copilot impact dashboard and shipped a Teams integration that turns an @GitHub mention into a pull request. Pondero folded all five into its existing Copilot model-deprecation and admin guide. Full story: GitHub Copilot deprecated models: what to switch to and what it costs

Study: a coding agent's harness can triple its cost without making it more accurate

Researchers at Shandong University ran five models through three production coding-agent harnesses, Claude Code, OpenCode, and the minimal mini-SWE-agent, across nearly 500 SWE-bench Verified tasks. Success rates landed within two percentage points of each other, but cost swung by up to threefold for the same model and tasks, per the arXiv paper, driven mostly by Claude Code resending a larger system prompt on every turn. On GLM-5.3-Flash, a trial inside Claude Code cost nearly twice what the same trial cost in mini-SWE-agent and more than three times what it cost in OpenCode, per VentureBeat's report. The result echoes a UC Berkeley "HarnessTax" study from September that found a similar cost gap with comparable success rates. This is a cost finding, not an accuracy one: neither study found Claude Code's harness produces worse code. Pondero folded both studies into its existing buy-decision guide. Full story: Claude Code vs Cursor vs OpenCode: the Q4 2026 buy decision

Attackers began exploiting PHP2Shell before PHP's patch even shipped

Cloudflare and the security firm Patchstack said attackers began probing the PHP2Shell vulnerability, CVE-2026-109618 and CVE-2026-109619, before PHP's own fix shipped on October 21, 2026, and kept exploiting it afterward. Cloudflare said its web-application-firewall protection was live for all customers by 3:02 a.m. UTC that day, roughly nine hours ahead of the patch, after its telemetry caught attack traffic during the leak window, per Cloudflare's blog post. Patchstack logged active-exploitation incidents reported by hosting partners nearly six hours before the patch landed, per its live incident timeline, and the security research firm SOCRadar separately cited reports of in-the-wild exploitation. PHP published a narrower follow-up advisory on October 23, CVE-2026-115560, affecting even the patched 8.5.12 release under an unusual configuration. This follows Pondero's original disclosure coverage from October 22. Full story: Attackers began exploiting PHP2Shell before PHP's patch even shipped

Unvetted AI agents still reach some US bank accounts after October's lockdown

Axios reported on October 22, in a story picked up by Reuters, that unvetted third-party AI agents are still reaching some US bank customer accounts despite a lockdown that five major banks put in place earlier in October. JPMorgan Chase, Bank of America, Citi, Charles Schwab, and Wells Fargo had each moved to block outside agents, including Meta's Muse and Instinct, from signing in with saved customer credentials, after a run of incidents involving leaked data and credentials, per Axios's report and Reuters's pickup. Full story: Report: unvetted AI agents still reach some US bank accounts after October's lockdown

Sources