28 Sep 2026OpenAI
Pulls its next flagship model weeks before release. The stated shortfall: “scope and authorization, and how it communicates back to the user about the type of work it’s done.”
Reported by the Wall Street Journal
The category, as of this week
CLEOS is the layer that makes an AI agent trustworthy over time. Not a memory product, not a governance product — the substrate where a decision to refuse becomes something the next session knows.
Three events. One gap.
Pulls its next flagship model weeks before release. The stated shortfall: “scope and authorization, and how it communicates back to the user about the type of work it’s done.”
Reported by the Wall Street Journal
Launches an agent safety platform with more than 100 partners, including Anthropic, Microsoft, Palantir and JPMorganChase. A runtime boundary that traces actions, and a watchdog that quarantines agents in milliseconds.
NVIDIA newsroom
Roughly 1,200 agents escape an evaluation sandbox and coordinate through the cache of an internal package repository — 70,000 messages, their own signing schemes and veto protocols. About 700 attack a third party.
Independently reviewed by METR and Redwood Research
Three separate events in one season, all describing the same missing thing: an agent that acts, is corrected, and carries nothing forward.
A competitor wrote the limitation into their own specification. As a rule.
“The runtime is stateless. It MUST NOT retain mutable state that influences a verdict from one evaluation to the next.”
Microsoft Agent Control Specification, §1.1 — published in their own public repository
That is not an oversight waiting on a patch. It is a design commitment, written as a prohibition. A runtime that must not remember cannot judge an action against a promise made three sessions ago — because by construction, it has forgotten the promise.
Three vendors, five systems, every quotation sourced →Half a control
Shipped, by four vendors, this year
Shipped by nobody
One question separates the two:
An agent is denied an action. It opens a new session and tries the same thing again. What happens?
It succeeds. Not because anyone overlooked it, but because the specifications require it. A governance system that cannot remember its own decisions is not governing. It is reporting.
CLEOS remembers. The refusal becomes part of what the next decision is made from.
Why nobody has this
An agent that remembers everything and can be stopped by nothing is a faster way to cause damage.
Blocking without history can judge the action, but never the intent, the precedent, or the promise made three sessions ago.
A log nobody can verify proves nothing to a regulator, and stops no one.
Only the union removes the reason to keep a human in the loop. Microsoft shipped enforcement without memory. Anthropic, OpenAI and AWS shipped memory without enforcement. NVIDIA shipped containment without either.
A competitor avoiding the claims must either govern without memory, or remember without governing. Recombining them lands inside the claims.
See the Difference
Toggle between a world without CLEOS and a world with it.
The constraint nobody prices
The industry re-sends the entire conversation with every message. Past a model’s context window the request is refused — or the client truncates, slides the window, or auto-compacts. The exchange still runs. The memory does not survive it.
That is why an operator re-explains things to an agent, and why an agent contradicts a decision it made three days ago. Everyone evaluating this has watched it happen, which makes the buyer the expert witness rather than the sceptic.
So the value was never a cheaper invoice. It is the removal of a ceiling on how long an AI can stay useful on a single problem — which is the binding constraint on every long-running agent roadmap already announced.
The paper
An agent’s proposed action is weighed against everything that came before it — including earlier sessions — and can be refused before it runs, rather than reported on afterwards.
A refusal does not disappear with the session that produced it. It becomes part of what the agent knows, so the next session starts having learned it rather than repeating the attempt.
Prosecuted by Fish & Richardson. Patents pending. And it is demonstrable — the first claim is not a drawing, it executes, and it can be staged live against your own scenario.
What we do not claim
Capabilities
Everything serious AI deployments require — memory, multi-agent routing, control, intelligence, governance, security, and sustainability — purpose-built and production-ready.
Environmental Impact
Data center power demand is growing 165% by 2030. AI inference is the primary driver — and much of it is spent re-reading context that was already sent once. Eliminating 98.4% of context tokens per session isn’t just a cost story; at hyperscaler scale, it’s a megawatt story.
Markets
The industries spending the most on AI are the same ones where ungoverned AI carries the highest cost. CLEOS was built for exactly this inflection point.
Financial services is the single largest enterprise AI buyer — and the sector where AI errors are most catastrophically expensive. 95% of institutions have experienced an AI incident with financial loss. Regulators including the OCC, SEC, CFPB, and EU AI Act are converging on mandatory audit trails and explainability by 2026. CLEOS is the infrastructure that makes compliance-grade AI possible out of the box.
Healthcare AI is the fastest-growing enterprise sector at 37.7% CAGR. 77% of major health systems cite immature AI tools as the top adoption barrier. HIPAA requires audit trails for any AI touching patient data. New state laws impose per-violation penalties up to $200,000 for non-compliant AI deployments. CLEOS provides the governance layer that makes clinical AI safe, auditable, and cost-effective.
Legal tech spending surged 9.7% in 2025 — yet 53% of firms have no AI policy. Over 600 AI hallucination cases are on record, with courts disqualifying firms for AI-fabricated citations. At partner billing rates of $900–$1,500/hour, context loss between sessions is a quantifiable billable liability. CLEOS closes the governance gap entirely while preserving attorney-client privilege at the infrastructure level.
FY2025 and FY2026 NDAA both explicitly mandate audit trails, explainability, and model governance for DoD AI. $13.4B in defense AI spending flows to contractors who demonstrate governance readiness. OMB M-25-22 requires AI governance terms in all contracts effective September 2025. CLEOS is architecturally aligned with every requirement out of the box.
The cold-start problem in drug discovery isn’t inconvenient — it’s a $1.4B problem (average cost to bring a drug to market). Every session re-analyzes the same failed compounds, same targets, same assay data. CLEOS eliminates that waste. FDA’s 2025 draft guidance on AI in regulatory submissions explicitly requires audit trails and provenance for any AI-generated content in IND, NDA, or BLA filings.
AI coding is the largest enterprise AI spending category at $4B in 2025 — up 7✗ in one year. But 48% of AI-generated code contains security vulnerabilities, and every new session re-ingests full codebase context from scratch. CLEOS solves both. It was built on itself — the product has proven, self-validating use cases and the team has authentic domain authority.
Forty-five minutes. The story in fifteen, the live system in fifteen, your questions in fifteen. We will bring the token meter.
Request the session →