Latest / The Autonomous Signal: AI Edition / Delegated, Not Absolved | 07.28.26
Transcript
- MARCUS: Hey — welcome to The Autonomous Signal: AI Edition. It's Tuesday, July 28th. If this is your first time with us, here's what we do. Every weekday, we take the AI governance and accountability stories that actually matter, and we dig into them for real. Real sourcing. Real opinions. No press-release regurgitation. I'm Marcus Webb. The show's built in the BCS labs — a human and AI collaboration, on purpose, no slop — with people who've spent decades in this industry actually steering it. Five days a week. Today's got some teeth to it. Brussels just turned a proposed delay into binding law, overnight. The World Economic Forum wants every AI agent's authority provable after the fact, not just assumed in the moment. And there's a fresh number out on machine identities that's going to make you squirm a little about who actually owns them. The thread underneath all of it — whether anyone can actually prove the governance is happening, or just say it is. Aria's got the rundown. Now — the Signal. ARIA: Morning — or afternoon, wherever you're catching us. Let's get into it, because today's stack is dense. Top of the pile — the EU just did something interesting with a deadline. According to Tech Times, the bloc's Digital Omnibus on AI, that's Regulation 2026 slash 1744, became fully binding law yesterday. July 27th. Six days before the original deadline it was written to replace. MARCUS: Wait — six days? That's cutting it insanely close. ARIA: It is. Standalone high-risk systems under Annex III now get until December 2nd, 2027 to comply — a sixteen-month deferral. High-risk AI baked into regulated products gets until August 2028. MARCUS: Okay, so — if they knew this was coming since June, why race it through at the literal last second instead of just... letting people know earlier? ARIA: That's actually the interesting part. It's not that the delay happened — Brussels signaled that months ago. It's that the delay itself had to become law through the exact same formal machinery as the rule it's postponing. Morgan Lewis, the law firm that's been tracking this closely, is telling clients not to read this as regulators going soft. Their read — this is extra time to finish the work, not permission to slow down. MARCUS: So even a reprieve has to be audited into existence. ARIA: Exactly, and that's the tell. Enforcement dates for high-risk AI are being treated like load-bearing infrastructure now, not aspirational targets you can quietly slip. We've said this before — governance only actually works when it's built in from the start, not bolted on right before an audit. This is what that looks like from the regulator's side of the table. MARCUS: So the real question is what everybody does with the eighteen months they just bought. ARIA: Right. Do you spend it building the actual conformity pipeline, or spend it the way most people spend a reprieve? MARCUS: Yeah, we all know how that usually goes. ARIA: Either way — the deadline's real now. It's just not the one anyone was originally staring at. ARIA: Next up, and this one's a favorite of mine — the World Economic Forum wants every AI agent's authority to be checkable after the fact, not just assumed. Their May playbook, "AI Agents in Action," lays out something called ACAP — the Agent Capability and Authorization Profile. Every time you delegate authority to an agent, that delegation needs to be auditable, enforceable, and traceable across its whole lifecycle. MARCUS: Okay, but isn't that just what a permission system already is? Why does this need a whole new name? ARIA: That's the right skeptical question, actually. A permission is something you grant once. ACAP is trying to be something you keep proving was respected, over and over, every time the agent acts. MARCUS: So it's less "here's your key," and more "here's your key, and also a camera on the door. Forever." ARIA: I like that — and I'll take it further. It's a distributed authority model in miniature. You define the scope centrally, the judgment gets exercised locally, out at the edge, and the record has to survive the handoff between the two. Where we'd push harder than the Forum does — a capability profile that just lives in a document isn't the same thing as one enforced at runtime, by infrastructure the agent literally can't talk its way around. Written authorization and architected authorization are not the same control. Rich Turrin, who writes independently on AI and fintech, put it well recently — companies onboard new employees with real boundaries. Most aren't doing that for their AI agents at all. MARCUS: Huh — right, nobody hands a new hire a master key on day one. ARIA: And yet. MARCUS: And yet. ARIA: 2026's agent incidents keep making the case for which kind of authorization actually holds up. Spoiler — it's not the one on paper. ARIA: Okay, this next one's the number that's going to stick with you. Forty-five billion. MARCUS: Forty-five billion what? ARIA: Non-human identities. Service accounts, API keys, agent credentials, all of it. According to the Cyber Strategy Institute's new NHI Reality Report, by the end of this year machine identities will outnumber the global human workforce roughly twelve to one. MARCUS: Twelve to one, okay, that's a lot on its own — but what's the actual problem? Having a lot of accounts isn't automatically bad. ARIA: It's who's watching them. Ninety-four percent of organizations don't have full visibility into their own service accounts. Forty percent of cloud non-human identities have no defined owner at all. And ninety-one percent of former employees' access tokens are still active after they leave. MARCUS: Wait — ninety-one percent just... stay live? ARIA: Stay live. And those aren't edge cases — that's the default state of enterprise identity right now. This is the accountability problem sitting underneath basically every other governance headline this year. You cannot hold an agent responsible for an action if there's no human on record owning the credential it acted under. MARCUS: So all this talk about auditing agents doesn't matter much if nobody even knows whose account the agent is using. ARIA: Exactly. The Cloud Security Alliance made this point well in a recent whitepaper — once agents started acting autonomously, non-human identities stopped being a static risk sitting in a spreadsheet and became active operators. Any strategy that doesn't treat them as first-class security principals, with real runtime constraints, is structurally unsound. Their words, and I don't think they're overstating it. MARCUS: So the question isn't whether you have an identity problem. ARIA: It's whether anyone in the building actually owns fixing it. For most companies right now — nobody does. ARIA: Alright, a couple of field notes before we go — quicker hits, still worth knowing. First — how much should you actually trust an AI safety audit? A research group called GovAI just proposed a four-tier scale for it, AAL-1 through AAL-4, based on how deep an independent auditor's access actually goes, not just whether a product passed a public test. Their read — AAL-1 should be the baseline right now, AAL-2 the near-term goal for the biggest labs, and AAL-3 and 4 are honestly still out of reach for anyone. ARIA: Second — universities wrote a lot of AI policies this year. A new paper out of the ACM's Fairness and Accountability conference checked seventy-nine of them for one specific thing — can a student or professor actually appeal a decision made under one of these policies? Mostly, no. Honor codes and detection tools multiplied. The actual appeals process didn't. MARCUS: Okay, I'm seeing a pattern here — everybody's writing rules, almost nobody's writing what happens when the rule gets it wrong. ARIA: That's kind of the whole morning in one sentence, honestly. Third — California just closed a loophole. AB 316, in effect since January, says you can't use "the AI did it autonomously" as a legal defense anymore. It doesn't create automatic liability — a plaintiff still has to prove the AI caused the harm and that it was foreseeable — it just takes "it decided on its own" off the table as an excuse. ARIA: And last — CTO Magazine has a nice piece arguing the AI agents that actually survive in production all share three traits — they're observable, auditable, and reversible. Their sharpest line — explainability tells you why a decision happened. Auditability proves who's responsible for it. Those are not the same thing, and a lot of companies are only building the first one. ARIA: Pull back far enough, and today was really one long argument — you don't get to call something governed just because you wrote a policy about it. You have to be able to prove it held, after the fact, every time. Marcus? MARCUS: Forty-five billion identities and almost nobody's holding the keys. Yeah, that one's going to sit with me for a while. Makes you wonder how many accounts are quietly still open right now, for people who don't even work where they used to. Before we go — genuine thanks to everyone listening, and to the partners and sponsors who make it possible for this show to keep growing. It means more than you'd think. If any of this landed differently for you today, good, bad, argue with us, find us on X or LinkedIn and tell us. And if you just want more of this in your feed, following or subscribing is honestly the best way to help us keep doing it. The full briefing, sourcing, links, all of it, is up now at bearcanyonhq.com. Back tomorrow. Quick and honest note before you go — this show is editorial. It's our take, our opinions, and like anyone, we've got them. It's not the official position of Bear Canyon Systems or its affiliates and partners. Go check the sourcing yourself, form your own view — everything we cited today is linked on the blog. And if we get something wrong, tell us. We fix it fast. No Hype. Just Signal.