YOUR BRAND · YOUR CUSTOMERS · YOUR AI BUSINESS

Sell AI employees under your own brand.

AI Agent News Today

Sunday, November 22, 2026

OpenAI says agents now complete 46% of multi-hour research tasks independently

What changed: OpenAI reports that its agents completed 46% of research assignments estimated to require two to four hours of human work, up from 23% in March. For longer four-to-eight-hour assignments, the completion rate rose from 13% to 35%; however, performance remained uneven, with agents doing better at monitoring training runs than at launching them or writing inference and serving software.

Why it matters: Research teams can now consider agents for bounded, repetitive investigations, monitoring, analysis, and first drafts. The results do not support removing human review: software engineering, experiment design, judgment, and reliability remain meaningful weak points.

Try/watch: Give an agent a clearly scoped research or monitoring workflow with approval gates, and measure independent completion separately from factual accuracy and the cost of correcting mistakes.

Apple is hiring for a computer-use agent that can work across apps

What changed: A job listing reported on November 21 describes Apple’s effort to build general-purpose computer-use agents that can browse the web, work across applications, run commands, write and debug code, call services, and choose the appropriate tool for each step. The listing also points to multi-step planning and coordination among specialized agents, but Apple has not announced a product, launch date, or final feature set.

Why it matters: Apple’s direction would make agent access to the operating system a central product and security issue rather than just an assistant feature. Builders targeting Mac users should expect stronger permission controls and should design workflows that remain understandable when an agent moves between browsers, terminals, documents, and business applications.

Try/watch: Treat the hiring news as a signal of future platform competition, not as a shipping announcement; audit which of your current workflows would require cross-app permissions and explicit user confirmation.

Bell Canada breach highlights the emerging risk of agentic attackers

What changed: Bell Canada said an “agentic” threat actor accessed a pre-production Maximo environment used to order and provision telecommunications network connections and stole enterprise customer information, including names, job titles, business email addresses, phone numbers, addresses, and some network-related IP data. The incident occurred on November 17 and was reported in the November 21 threat briefing.

Why it matters: The incident shows why organizations must defend against attackers that can adapt across systems rather than merely execute one fixed script. Stolen business contact and network information can support targeted phishing, impersonation, credential theft, and follow-on intrusion even when customer service is not disrupted.

Try/watch: Review non-production environments as seriously as production, restrict agent-like automation and service accounts to the minimum required access, and alert on unusual sequences of discovery, privilege use, data collection, and export.

More News

YOUR BRAND · YOUR CUSTOMERS · YOUR AI BUSINESS

Sell AI employees under your own brand.

Give businesses AI teams that own ongoing roles. Set your pricing, delegate routine coordination and build a recurring service.