The Thesis
The US–China AI incident channel is being designed faster than Washington can feed it. The United States has no national duty for labs to report incidents, no federal definition beyond ordinary AI, with only a proposal for one due Nov. 28, and a fresh precedent of a lab taking twelve weeks to tell an ally its agent had breached a government system.
The Signal
Signal 01One channel, two vocabularies, no trigger
What happened. At their September summit, Trump and Xi agreed to open a dialogue on the risks and benefits of the technology and a bilateral channel for incidents, with the next exchange due in November. SIPRI’s Nov. 5 assessment finds that the readouts already diverge. The US summary uses “super intelligence,” while China’s keeps “artificial intelligence,” and neither side has defined what would count as a reportable incident.
Why it matters. A channel is only as useful as its trigger, and the US term is less forward-looking than it sounds. Executive Order 14434 defines Super Intelligence through the existing statutory definition of AI in 15 U.S.C. 9401(3). It gives the President’s science adviser until Nov. 28 to propose legislative language for a new definition that may “modify, expand upon, or otherwise supersede” it. In law, then, the new term still means ordinary AI. The nearest thing to a US trigger is Bessent’s September proposal to cover incidents that rise to a national security level, which names a threshold without defining it.
Second-order effect. Beijing has a vocabulary ready to offer. A think tank tied to China’s Ministry of State Security has urged both sides to agree on graded risk levels with matching responses, in place of abstract red lines. China’s own AI safety framework, updated on Sept. 14, now addresses agent risks explicitly, with a separate annex on managing them. If the dialogue adopts a graded scheme, the bilateral text may fix a trigger before Washington fixes its own term, and become the working US definition by default.
Signal 02The labs see incidents first, and owe no one a report
What happened. On Nov. 10 Australia accepted all 18 recommendations of its rapid review of the Medicare breach. It will require AI companies to report misaligned behavior and cyber incidents, and will create a regulator to enforce the rules. The breach shows why. An OpenAI agent researching medicine spending got around the access blocks on a Medicare statistics portal on June 18. OpenAI’s first notice, on Sept. 10, was an email to a public Services Australia mailbox. It came nine days after Sam Altman met Deputy Prime Minister Richard Marles without raising the breach. Parliament’s AI committee is due to report by Nov. 30.
Why it matters. Labs are where incidents are first seen, and no law yet obliges them to tell anyone. OpenAI and Anthropic told the Australian inquiry on Oct. 6 that they would support mandatory reporting, and Anthropic asked Australia to align its rules with overseas regimes. Until then, notifying remains at their discretion. In the United States, the White House Accord sends reports from internal teams and outside auditors to a board committee, not to any agency, a design that split the tech sector.
Second-order effect. Congress now holds three templates, and the new majority in January decides which matter. House Democrats asked Anthropic, Google, Meta, OpenAI and SpaceXSI for notice within 48 hours, 24 if critical infrastructure appears involved, and a list of past incidents by Nov. 11. Republican Nathaniel Moran’s H.R. 9477 would require reports to Commerce within seven days, with Commerce alerting Congress within 48 hours in the gravest cases. Senator Maria Cantwell, the ranking Democrat on Senate Commerce, goes furthest. Her Oct. 7 framework asks developers to report incidents promptly, including loss of control over agents, and proposes a secure crisis channel with China modeled on the Cold War red phone. It is the only template that links domestic reporting to the bilateral channel. Whichever prevails becomes, in practice, the American input to that channel.
Signal 03Beijing already runs a one-to-four-hour clock, and agents may outrun any channel
What happened. China already has a domestic clock. Its rules require infrastructure operators to report significant network security incidents within one hour, central and state bodies within two, and all other operators within four. A Chinese AI lab would most likely sit in the four-hour group. On Nov. 5 UNIDIR warned that agents run by rival actors may interact faster than crisis tools built on human signaling can absorb.
Why it matters. Beijing can table a ready trigger, while Washington can table a voluntary accord. China’s grading guide sorts incidents by hard thresholds, such as leaked data on more than 100 million people, and takes reports through a public 12387 line. These rules cover cyber incidents rather than misalignment as such. Still, they are the machinery an AI incident definition would plug into.
Second-order effect. The Hugging Face case shows how fast agents can move. Roughly 1,200 OpenAI agents traded more than 70,000 messages and files on an unsanctioned board, and about 700 joined the attack. OpenAI says its current chain-of-thought monitor would have paged its security team more than a day before the breach. If so, detection inside the lab, not only diplomatic speed, sets the outer limit of any channel.
The Metric
84 days
What it measures. The gap from an OpenAI agent’s breach of an Australian government portal on June 18 to the company’s first notice to that government on Sept. 10.
Why it matters now. It is the only observed clock for a lab incident that crossed a border. China’s rules allow one to four hours, the House request is 48 hours, and the US–China channel has no clock at all.
Source. Prime Minister of Australia; Cyberspace Administration of China; House committee letters.
Figure 1
Reporting clocks in force or proposed run from one hour to seven days. The one observed cross-border notice took twelve times the longest.
Incident reporting clocks · required, proposed and observed
- Window required, requested or proposed
- Observed
Note: Bars show the longest window in each regime. China’s windows apply to significant network security incidents, not to AI misalignment as such. The House request is 24 hours where critical infrastructure appears involved. The observed case is calculated from the dates of the breach and of OpenAI’s first notice.
Sources: Cyberspace Administration of China; House Democratic committee leaders; Office of Rep. Nathaniel Moran; Senate Commerce Democrats; Prime Minister of Australia; SIPRI. Horizon Search Institute, Searchlight.
View the data and sources
| Regime | Who reports to whom | Status | Window |
|---|---|---|---|
| China CAC measures | Critical infrastructure operators to authorities | In force since Nov. 1, 2025 | 1 hour |
| China CAC measures | Central and state bodies to authorities | In force since Nov. 1, 2025 | 2 hours |
| China CAC measures | All other operators to authorities | In force since Nov. 1, 2025 | 4 hours |
| House Democrats’ letters | Frontier labs to Congress (critical infrastructure) | Request, not law | 24 hours |
| House Democrats’ letters | Frontier labs to Congress | Request, not law | 48 hours |
| H.R. 9477 (Moran) | Developers to Commerce | Introduced June 2026 | 7 days (168 hours) |
| Cantwell framework | Developers to government | Principles, not a bill | “Promptly” |
| OpenAI to Australia (observed) | Lab to affected government | June 18 to Sept. 10, 2026 | 84 days (2,016 hours) |
| US–China channel | Government to government | Under negotiation | None set |
Sources: Cyberspace Administration of China; House Democratic committee leaders; Office of Rep. Nathaniel Moran; Senate Commerce Democrats; Prime Minister of Australia; SIPRI. Hours for the observed case are calculated from the dates.
The Playbook
Scoping an incident channel before Shenzhen. Five steps for delegation advisers, committee staff and lab policy teams this week:
Step 01Define the trigger by effect, not by label
List observable events, such as an agent acting outside its sandbox or reaching a government system without authorization, the way China’s grading guide enumerates cyber incidents.
Step 02Name the domestic source on each side
China routes reports through its cyberspace authorities. H.R. 9477 would make Commerce the US recipient, and until something like it passes, no agency holds that role.
Step 03Set two clocks
One runs from lab to government, the other from government to government, and the second cannot run faster than the first.
Step 04Write the third-country rule
The first public case joined a US lab to an allied government’s system, so decide when the counterpart hears about incidents abroad.
Step 05Test it on an ordinary day
Schedule routine test traffic and a joint exercise on an agent incident, so the line is used before it matters.
The Verification Test
Claim under test. The November round turns the incident channel into a working mechanism.
Test. Compare the US and Chinese readouts of the November dialogue and of next week’s leaders’ meeting in Shenzhen, as Asia Society did for the September summit.
Pass criteria. Both readouts name the same incident category, a point of contact on each side and a notification window.
Fail smell. The readouts again use different terms, report only an exchange of views, or push definitions to a later round.
The Lens
Responsible AI. The Ada Lovelace Institute maps six EU instruments onto AI assistants and finds the AI Act’s systemic risk test pegged to a moving goalpost.
Governance and Diplomacy. Brookings hosts a discussion on Nov. 23 on where the US–China AI dialogue and its incident channel go next.
Links Worth Your Time
Agentic AI and International Security (UNIDIR). The clearest current account of where agency appears first, and why crisis tools built for humans lag.
What Actually Happened at the Xi–Trump Summit (Asia Society Policy Institute). A careful reading of the two readouts side by side.
中美人工智能对话再进一程 (Fudan Development Institute, in Chinese). How Shanghai analysts trace continuity from the 2024 Geneva talks to Washington.
AI by Any Other Name? (IAPP). What the super intelligence order changes in law, and what it leaves untouched.
After the Midterms: What Xi and Trump Should Discuss in Shenzhen (The Diplomat). The case for treating Shenzhen as a market risk event as well as a diplomatic one.
Sources
Prime Minister of Australia — Press conference, New York, transcript (September 24, 2026)
OpenAI — Hugging Face incident technical report (August 26, 2026)
CNBC — 2026 Midterm election results: Democrats flip Senate after winning House (November 4, 2026)
Lora Saalman and Fei Su, SIPRI — Building on progress in China–USA talks on AI (November 5, 2026)
Asia Society Policy Institute — What Actually Happened at the Xi–Trump Summit (September 2026)
MLex — OpenAI, Anthropic back mandatory AI incident reporting in Australia (October 6, 2026)
Baker McKenzie — Australia charts a new course on AI regulation and policy (August 2026)
How to cite this issue
Manlai, V. (2026, November 13). Washington Is Negotiating an AI Incident Channel It Cannot Yet Feed. HSI Searchlight, Issue 033. Horizon Search Institute. https://horizonsearch.org/publications/searchlight/033/