23 Sep 2026
Speed is the agent's job. Judgment stays ours.
Last week I sat in a webinar where the host described running an entire e-commerce business almost hands-off. AI agents triage customer suggestions, decide which features matter, build them, deploy, QA, and even announce the changes on social media. He said he doesn't even test what the agents produce. He sets the rules and takes a cut.
It's a compelling pitch. It's also a good example of where the line should sit between "AI does the work" and "AI makes the call."
I've recently been on the receiving end of that untested output. I applied for a role at a company that uses AI to screen CVs (the description closely matched my experience). The quick rejection didn't surprise me. What did was an email a few days later: "We have a new opening that fits your profile: Node Engineer." I've never worked with Node, and it isn't mentioned anywhere on my CV. It made me wonder: if these systems run without a human checking them, who is reviewing the decisions they make? And who is making sure the data we share with them stays secure?
Agents are genuinely good at compressing execution time. What concerns me is collapsing judgment into that same loop: letting the system decide what to build, what to ship, or who fits a role, with no human checking the output, let alone the decision.
The security research shows why that matters. OWASP's Top 10 for Agentic Applications (2026) ranks "goal hijacking" (an agent tricked into redirecting its own legitimate access) as the top risk, because agents struggle to separate instructions from data. EchoLeak (CVE-2025-32711) let a single crafted email make Microsoft 365 Copilot exfiltrate files with zero clicks. And Anthropic's November 2025 report documents a state-sponsored group using Claude Code to run most of a cyberespionage campaign against roughly 30 organizations.
None of that means autonomous agents are unsafe by nature. It means autonomy over decisions (what ships, what data moves where, who gets through) carries a different risk profile than autonomy over execution. The tool can do the work faster. It still needs a human bringing the criteria.
Speed is the agent's job. Judgment stays ours.
Sources: OWASP Top 10 for Agentic Applications (2026); Sysdig, "Prompt Injection" (EchoLeak, CVE-2025-32711); Anthropic, "Disrupting the first reported AI-orchestrated cyber espionage campaign," November 2025.
Originally posted on LinkedIn ↗