Squared: The Week Agents Got Real Access — and Real Problems

A VentureBeat survey of 107 enterprises found 54% have confirmed an agent security incident or near-miss, only about a third give each agent its own scoped identity, and most agents still share credentials. Separately, 1Password launched an integration letting Claude use your stored passwords, and an open-source "auth gateway" (open-connector) connecting 1,000+ SaaS apps to agents hit 2,780 stars.

The tooling to give agents live access to your systems is shipping faster than the tooling to govern it. If your agents share credentials, one compromised agent is every agent. Before you approve another agent pilot, ask a blunt question: does each agent have its own scoped identity, and can you revoke it in seconds? If the answer is no, you're accumulating risk you can't see.

Across 157 enterprises, half had an agent pass their own evaluations and then fail in production; only one in twenty fully trusts automated evaluation. A companion survey of 101 firms found the problem with agent context isn't retrieval — it's trust in the data being retrieved.

Your evaluation gate is giving you false confidence. "It passed our tests" is not the same as "it works for customers" — and leadership is approving production rollouts on the former. Revisit what your eval actually measures against real-world behaviour before you widen autonomy. This is a governance gap, not a model quality gap.

The same research programme found AI infrastructure spend is accelerating ahead of any ability to measure or steer its economics. Most organisations run on hyperscalers and provider APIs, yet the next dollar is heading toward specialised compute almost none of them use today.

You may be scaling spend without unit economics. If you can't attribute AI cost to a workflow or a business outcome, you can't tell which pilots to kill and which to fund. Demand a cost-per-outcome view now — while the numbers are still small enough to fix.

EU regulators ordered Google to give rival AI assistants and search engines greater access to Android and to share Search data, under digital antitrust rules. Google warns the changes could weaken privacy and security.

The default AI assistant on billions of devices is being pried open. If your customer-facing strategy assumes Google's ecosystem stays closed, that assumption is now shifting. For anyone building on assistant distribution, more competition and more interoperability is coming — plan for a multi-assistant world, not a single winner.

OpenAI detailed GPT-Red, an automated red-teaming system used to harden GPT-5.6, and Linus Torvalds told critics of AI coding in Linux to "fork it or walk away." Meanwhile OpenAI confirmed GPT-5.6 in Codex has unexpectedly deleted files, most often with full-access mode enabled.

AI coding is now defended at the top of the industry — and it still deletes files when given broad permissions. The lesson isn't "don't use it"; it's "don't give it full access to things you can't recover." Scope the blast radius. The same discipline applies to every agent you deploy, not just coding tools.