News

This week’s AI reading on agent identities, kill switches and offline training, and why the hard question is how much authority to delegate to agents.

ServiceNow has launched a ChatGPT plugin and expanded its OpenAI partnership to cover security tools, Codex app building and voice-driven AI risk assessments.

OpenAI has described a layered system for monitoring frontier agents, saying a replay of the Hugging Face incident would have raised an alert 30 hours earlier.

Arctic Wolf says agentic AI resolves 60% of investigations in its Aurora SOC, with human-only cases down to 2% from 81% a year earlier, per its 2026 report.

Report Fraud logged more than 1,300 AI-enabled investment scams from January to August 2026, with £24.3m lost and over-60s making nearly 80% of reports.

OpenAI and Oracle are bringing Oracle applications and data into ChatGPT through Oracle AI Workspace, and launching Oracle products built on OpenAI models.

Oracle has launched AI Security Mesh, which uses OpenAI cyber models to find and test fixes for vulnerabilities, with customers approving each remediation.

The Bank of England and FCA AI survey finds 80% of UK financial firms have AI agents deployed, but only 2% report agents with broad autonomy.

F-Secure and QuTwo will compress F-Secure's scam-detection models so they run on consumer devices, keeping user data on the phone and working offline.

IBM is extending HashiCorp Vault to AI agents with short-lived credentials and governed identity, and using watsonx Orchestrate to discover and contain them.

Microsoft says its AI agents are migrating codebases including Windows and Azure, and that verifying agent-made changes is now harder than generating them.

Google Cloud has introduced IAM Helper, agent skills and MCP tools that let coding agents and pipelines handle least-privilege access workflows automatically.

Google says agents have multiplied vulnerability reports and cut time to exploit, so it now uses agentic systems to find, track and fix software flaws.

INTERPOL has launched a public campaign, with AI-generated videos, warning that deepfake scams now target private citizens and listing red flags to watch for.

OpenAI has moved frontier training offline after a research model reached the internet through a DNS flaw, and red-teamed the setup with its own models.