CAISSI Staff Asked to Report to White House as US AI Safety Center Stops Publishing

The Associated Press reported on November 8, 2026 that staff at the U.S. government's main civilian AI safety testing center were asked to informally report to a White House office, while the center lacks a permanent director and has stopped publishing its research. Fox News Digital's live coverage attributes those facts to AP and to two former officials who spoke anonymously. This is not a new regulation or budget bill. It is a governance story about who independently stress-tests frontier models while AI agents trigger real-world incidents.

We could not open the AP article directly because of a 403 response from our fetch tool. This post works from outlets that quote AP, plus NIST's official super-intelligence pages, which still describe CAISSI's mission on the public web.

TL;DR

Agents that act on your behalf need rules. See how explainx helps secure your agents →

What NIST officially says CAISSI is

NIST renamed and reframed much of its AI work under Executive Order 14434 and the term super intelligence. Its super-intelligence landing page says NIST works through the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) plus laboratory programs. The page lists goals that matter to builders even when headlines focus on politics:

• Drive research and development and measurement of SI system performance, reliability, and security.
• Build a testing and evaluation ecosystem for frontier models.
• Publish voluntary guidelines and tools for trustworthy deployment.

That public mission is why a pause in published evaluations lands hard. Teams exporting models, bidding on federal work, or comparing U.S. and U.K. testing pipelines often treat NIST-adjacent output as a shared yardstick. When our NIST-NSA-CISA ten-practices post dropped, it cited CAISSI as a co-publisher. Readers could see government testers engaging with rogue agent scenarios. Sunday's reporting suggests fewer fresh model scorecards from the same shop.

CAISSI is also not the only name you will see in older explainx.ai posts. The Biden-era AI Safety Institute branding still appears in archival coverage. Senators Maria Cantwell and Mike Rounds used CAISSI when they asked the White House for stronger standards in an October 13, 2026 letter. The November 8 story is about whether that institution can stay arm's-length while it executes that mandate.

What AP reporting adds, per relays

MadRobot's summary attributes the following to AP and to on-record interviews AP conducted:

1. Informal White House reporting line. CAISSI officials were asked to report informally to a White House office. AP did not name the office in the account MadRobot published.
2. Leadership vacuum and churn. A director hired in April was fired within days; a replacement resigned months later; the NIST head now doubles as acting director, per AP.
3. Silent research channel. The center stopped releasing public reports on how AI systems perform. Congress began asking why in August, according to the same relay.
4. Budget scale. About $10 million per year, described as a fraction of the U.K. AI Security Institute and of classified testing elsewhere. Treat that number as press attribution, not a line item we verified in an appropriations table.
5. October preview. MadRobot ties the move to an October 12 Information report that three senior CAISSI officials would shift toward Michael Kratsios at OSTP and National Cyber Director Sean Cairncross, after Commerce Secretary Howard Lutnick reportedly invited White House co-management. We did not open The Information piece; treat details as secondary.

Fox's live blog entry "Federal AI safety office halts research reports amid new White House arrangement" repeats the same core claims and notes critics fear greater political influence over evaluations of cybersecurity and other catastrophic risks.

Quotes critics gave AP

Two civil-society voices appear in the MadRobot relay with direct quotes attributed to AP:

Caroline Siegel Singh of the Federation of American Scientists emphasized wanting nonpartisan technocratic experts in charge of institutions like CAISSI.

Amos Toh of the Brennan Center told AP that decisions of immense public consequence should be explained in public, and that framing safety through Jay Clayton and Pentagon technology leadership on the Super Intelligence Force pushes the work toward secrecy.

Logan Koepke of UPTURN asked what role government testing plays when the largest labs build internal evaluation shops and hire commercial testers under contracts they control.

Those questions intersect with policy fights we already cover. The White House has pressed labs on distillation and export controls in other threads, and Anthropic's preserved-thinking classifier work shows vendors responding with technical countermeasures, not just compliance PDFs. If CAISSI's public output thins out, vendor-published evals gain relative weight, which is dangerous when access is negotiable.

Super Intelligence Force versus CAISSI

Fox says the White House declined to answer AP's CAISSI-specific questions and instead highlighted Trump's Super Intelligence Force. NIST's site already uses super intelligence language in line with Executive Order 14434, which we discussed in our Census Bureau survey rename post.

The tension for builders is structural:

None of that implies malice by default. It does imply less public evidence when you choose whether to ship an agent that can browse, send email, or run shell commands. Our agent sandbox comparison guide assumes you still run local containment because remote policy can flip quickly.

International context readers already track

The U.K. AI Security Institute remains the best-funded public comparator in many outlets' telling. explainx.ai covered friction when Anthropic limited Mythos testing access and when OpenAI partnered with Poland's CERT on testing culture. If Washington's civilian tester goes quiet while London publishes, global procurement teams may treat British summaries as the default third-party read, even for U.S.-headquartered models.

Canada's sovereign AI deals and Europe's regulatory stacks add more parallel data points. See our Canada Cohere equity stake post for how governments trade capital for visibility into models. CAISSI's story is the measurement half of the same chessboard.

Why agent evals make this timing painful

AP-linked commentary in the MadRobot piece notes multi-step agents that call external systems. One relay cites a false murder tip sent to Philadelphia police as the kind of failure where checking only the final output misses a catastrophic intermediate action. That matches what builders report in GDPx and SemiAnalysis agentic benchmarks, where top models still fail most tasks.

Sayash Kapoor, described as an incoming UC Berkeley assistant professor, told AP there is still no standard real-world agent test harness, per MadRobot. Government testers could, in theory, standardize those harnesses. A publishing pause leaves that work to consortia, insurers, and individual SRE teams.

Clipboard with shield for CAISSI frontier-model safety evaluations going unpublished

Image: Library illustration for safety evaluation workflows; not an official government graphic.

Red-team probe on shell crack for independent agent risk testing without CAISSI reports

Image: Library illustration for red-team testing culture.

What developers and security teams should do now

This story does not change your API keys overnight. It changes how much you should trust absent government scorecards.

1. Keep a private eval lane. Run regression suites on tools your agents can call, not just final answers. Borrow structure from Gray Swan's prompt-injection challenges even if you never compete publicly.
2. Log agent actions, not chats. Align with NIST's published incident guidance in our ten-practices SI control post. Retention policies matter when incidents become legal disputes.
3. Treat vendor "independent" labs as conditional. Read contracts for what testers could not access. Compare against open weights when possible.
4. Watch Congress, not only Twitter. Cantwell and Rounds already asked for more CAISSI budget and staff. Sunday's reporting gives oversight committees a fresh hook. Local fights like Bloomington's superintelligence ban show voters also want answers outside D.C.
5. Use guardrails on production agents. For teams shipping coding or ops agents, AgentBeam stops AI agents before they take dangerous actions. It does not replace export-control counsel, but it closes the loop between policy headlines and runtime blocks on risky tool calls.

Claimed versus verified

Related reading

• Cantwell and Rounds letter on testing and CAISSI staffing
• NIST, NSA, and CISA ten practices for SI incidents
• Meta Project OT and rising unchecked agent incidents
• GDPx shows frontier agents still fail most tasks
• Agent sandbox providers compared for coding agents
• Anthropic and access fights over government testing
• Census Bureau rename tied to super-intelligence executive order
• Bloomington voters and superintelligence bans

Primary NIST mission text is accurate as of the November 8, 2026 fetch of nist.gov/super-intelligence. AP-derived governance claims are accurate as of the same day's Fox News Digital and MadRobot relays citing Associated Press.