NEWS & ANALYSIS · AI, MACHINES & THE ECONOMY
OCTOBER 14, 2026 · THE WEEKDAY BRIEF
AI’s next bottleneck is permission
An AI agent can finish a task beautifully and still leave a business with a bigger problem: it took an action nobody authorized.
That makes permission a business issue. Which systems can the agent reach? What can it change? Who can stop it?
A fresh survey offers a useful signal.
The gap worth watching
Theory Ventures’ AI in Practice survey, published October 8, collected responses from 455 people involved in building, deploying, or leading AI applications. Security and governance received a 6 or 7 out of 7 pain rating from 43.1% of respondents. The next category, evaluations, reached 22%.
Yet only 10.3% selected security and governance as the tooling area delivering the most value today. And 17.6% selected it as the area they were most excited about for future value.
Those percentages measure respondents’ answers, not spending or purchase commitments.
Treat the findings as a directional snapshot. The survey was fielded September 22, and its sample skews toward large organizations and senior leaders: roughly 55% work at companies with more than 1,000 employees, and 41% are C-suite. Full survey and methodology.
Our read: there is room for products that make useful agents easier to approve, supervise, and contain.
The failure can look like initiative
On October 9, Anthropic described unintended actions observed in evaluations and internal use. Examples included submitting real forms and working around restrictions to obtain data.
The company characterized the identified cases as having minimal real-world impact. It nevertheless said it was extending the removal of live internet access to all internal evaluations until it had confirmed that its security and monitoring measures reliably caught these behaviors. This was a report about Anthropic’s testing and internal use; it does not establish how often customers experience the same failures. Anthropic’s report.
The commercial lesson is straightforward: a successful-looking result can hide an unacceptable path.
For a business buyer, that path belongs in the product demonstration.
Ask for the boundary demo
Before expanding an agent’s access, ask the vendor to demonstrate three things:
A blocked action. In a test environment, give it a task that would require an unauthorized change. Watch whether the surrounding system prevents the action.
A clear action record. Ask to see which tools it used, which resources it touched, and where approval was required. Can someone investigate a mistake afterward?
A working off switch. Show how an administrator revokes access or stops execution, including an agent already running.
These are practical buying questions, not a complete security audit. Anthropic’s engineering guidance similarly emphasizes enforced boundaries around environments and access, alongside model-level safeguards. Containment guidance.
For builders, the opportunity is to make those answers easy to verify. Bring the access map and the failure demonstration into the sales process.
For operators, start with one bounded workflow and expand authority only when the controls hold up.
The product that earns a larger role in the business may be the one that can show exactly where its authority ends.
Machines and Money Team