Skip to content
OriginalSecurity

Microsoft's fix for un-auditable AI agents is a standards process with no deadline

At Ignite, Microsoft admitted companies can't verify what their AI agents do — then announced a standards effort to fix it. The catch: there is no spec, no log format, and no deadline yet.

Cyanotype photogram: white silhouettes of a key, a coiled cable, a magnifying glass and a folded ledger on deep Prussian blue paper, with a faint ghost outline of a second key beside them
A cyanotype of the tools of the trade — key, cable, ledger, magnifying glass — and one ghost key: the action no log captured.
On this page

In brief

  • Microsoft announced the AI Transparency Initiative at Ignite on Nov 17, with more than 40 organizations, to build open standards for seeing, verifying and controlling AI agent activity.
  • The pitch is an evidence gap: existing controls watch the activity around a model but cannot reconstruct what happened between a request and an outcome.
  • The announcement names no specification, log format or deadline — and NIST's SI Agent Standards Initiative, launched in February, has so far produced voluntary guidelines, not standards.
  • NIST's own red-teaming found novel agent-hijacking attacks succeeded 81% of the time, against 11% for the strongest known baseline — the risk the standards are meant to close.
  • Microsoft will build the work into its AI models, Foundry and Copilot, and says evidence should stay with customers when models and providers change.

Microsoft has a name for the problem at the centre of its Ignite keynote on Tuesday: the evidence gap. Companies can watch most of what happens around an AI model — the request that went in, the tool calls that came out — but they cannot reconstruct what happened between the two, or independently verify the model's own account of its behaviour. When an agent with access to data, tools and credentials does something consequential, the organisation is left taking its word for it.

The fix Microsoft announced, the AI Transparency Initiative, is a promise to close that gap with open standards — built with more than 40 organisations, developed alongside NIST's agent-standards programme, and eventually wired into Microsoft's own AI models, its Foundry platform and Copilot. The diagnosis is right and the moment is right. There is one catch, and it is not small: the announcement contains no specification, no log format, no deadline and no ship date. The standard does not exist yet. What exists is a commitment to write one.

The gap Microsoft is naming

The initiative organises its answer around three words: see, verify, control. See what the AI was asked and authorised to do, which model and runtime ran it, what data and tools it touched, and what actions it took. Verify those actions independently, because, in Microsoft's own words, "agents cannot be the sole source of truth about their own behavior." Control what the AI is permitted to do, and keep the evidence when models and providers change.

"Blaming the AI isn't an answer, and 'trust the model' isn't a control," the release says. Accountability, it adds, stays with the organisation — leaders answer to customers, employees, boards, auditors and regulators. The problem is that agents now carry the risk profile of a privileged insider while operating at machine speed and scale, and the controls built for human insiders do not fit them. A 1Password survey published in July found AI agents regularly access unapproved data at 41 percent of the organisations that responded.

At the same event Microsoft said Agent 365, its control plane for governing agents, now discovers 60 agent types — including Claude Code, OpenAI Codex and GitHub Copilot — and that Microsoft Perception, its system for orchestrating security agents, is generally available. Those products watch the agents. The Transparency Initiative is aimed at the part they cannot see: what happens inside the model itself, between the request and the outcome. That is the part no vendor dashboard covers today, including Microsoft's.

What NIST has actually produced

The initiative's first leg is an expanded Microsoft role in NIST's SI Agent Standards Initiative, the federal programme meant to produce the open standards this work would be built on. The programme launched on February 17, 2026, as the AI Agent Standards Initiative, and its page was last updated on November 9. It has three pillars: industry-led standards, community-led open protocols, and research into agent authentication and identity. It ran a request for information on agent security that closed on March 9, and published a concept paper on agent identity and authorisation that was open for comment through April 2.

Nine months in, what the programme has produced is voluntary guidelines, gap analyses and concept papers — not a standard. That is consensus speed, and a measure of how far the work has to go. Microsoft's announcement does not say which of its workstreams it will join, or how its work maps onto them.

The "SI" in SI Agent Standards Initiative stands for "Super Intelligence." The label is new: the February announcement called it the AI Agent Standards Initiative. It changed because Executive Order 14434, signed on September 29, 2026, directs federal agencies to use "Super Intelligence" and "SI" in place of "artificial intelligence" and "AI" in official communications. The order's own definition of the new term is the existing statutory definition of AI, word for word. So the federal standards programme for agents has been renamed, but the thing it standardises has not changed — and neither has the fact that its standards do not yet exist.

Why the gap matters now

NIST's own red-teaming, published in January 2025, tested whether an AI agent could be hijacked by malicious instructions hidden in data it ingests. The strongest previously known attack succeeded 11 percent of the time. The strongest new attack the red team developed succeeded 81 percent. When attacks were repeated 25 times, the average success rate rose from 57 percent to 80 percent. One clean test run proves very little — and a compromised agent's own logs are the last thing to trust about whether it was compromised.

Microsoft's own security blog, published alongside the announcement, supplies the scale of the surrounding problem. More than 96,000 vulnerabilities have been published to the CVE programme in 2026, more than double 2024's roughly 40,000, driven in part by AI-assisted discovery. When PHP disclosed two critical remote code execution flaws this fall, tens of millions of internet-facing services were running affected versions; nearly two weeks later, about 29 million remained vulnerable. Researchers had a working, weaponised exploit within roughly a day of disclosure. Only 26 percent of the vulnerabilities in CISA's Known Exploited Vulnerabilities catalogue were remediated in 2025.

This is the week that context lands in. Google's threat researchers reported the first fully automated vulnerability discovery and exploitation pipeline they have seen running with no human in the loop — a framework that reads patches, drafts exploits and fires them. An attacker poisoned a single Optimizely snippet and served malware from 178 organisations' websites, because the trusted vendor's code was the attack surface. On the defensive side, OpenAI announced NATO's institutions are running its cyber models at up to 300 billion tokens a month. The agents are already deployed, on both sides, at industrial volume. The audit trail is a future standard.

What would count as progress

None of this makes the initiative a bad idea. The diagnosis is precise, the coalition is real — Accenture, Adobe, KPMG, Mastercard and Volvo Cars are named participants — and the design principle is right: evidence should be checkable against independently observed records, and it should stay with the customer when models and providers change, rather than living inside one vendor's dashboard. A common evidence format for agent activity is genuinely missing, and Microsoft is well placed to start building it: it sells the agents and the tools that govern them.

But a standards process is judged by its deliverables, and this one has none yet. The things to watch for are concrete: a published evidence schema, a log format that independent tools can read, a dated milestone, and adoption by someone other than Microsoft. Until those appear, the honest status of the AI Transparency Initiative is that Microsoft has named the problem exactly and assembled the coalition to solve it — and the solution is pending.

The announcement's best line is also its hardest test. "The more agency we give AI, the more transparency we should expect in return." The agency has already been given. The transparency is still a sign-up page.

More from grep.tips

All articles