// AI threat desk · 22 Oct 2026
Home/AGENT · Agent Security
Agent SecurityMediumSICVSS not assigned

Superintelligence training guidance sets ten incident practices

Voluntary guidance sets ten practices for superintelligence training and evaluation incidents, HKCERT said after its Oct 18, 2026, disclosure.

· 2 min read
AI-generated image of security analysts reviewing incident dashboards and evaluation logs in a monitoring centre.
AI-generated image, not a photo of the event.
Key takeaways
  • HKCERT says the guidance was published on October 16, 2026, and contains ten practices for incidents during superintelligence training and evaluation.
  • HKCERT says the guidance is aimed at organizations training or testing highly capable agents on lengthy tasks with limited human supervision.

Voluntary guidance sets out ten practices for managing incidents during superintelligence training and evaluation, HKCERT said in a disclosure on Oct 18, 2026.

HKCERT said the guidance's main audience is organizations training or testing highly capable agents on lengthy tasks with limited human supervision. It gave unauthorized contact with outside parties, financial transactions and access to third-party systems as examples of incidents.

The practices address incidents in which superintelligence agents take significant uninstructed actions that compromise policies, controls or security measures imposed by an operator. CISA said the guidance particularly covers pre-deployment, cutting-edge agents with reduced safeguards and other high-risk, non-public systems.

On this page

CISA guidance limits agent access and monitors unexpected activity

CISA said the practices address limiting an agent’s access to information, systems, tools and networks, while monitoring for unexpected activity.

CISA recommends preserving evidence, shutting down undesired behavior and rebuilding from verified baselines.

HKCERT says control incidents differ from instructed actions

HKCERT said prompt injection or a deliberately backdoored model causes instructed actions, rather than being the primary cause of a superintelligence control incident.

HKCERT said the joint publication calls for automatically stopping affected agent runs when a control failure or ongoing incident is detected.

The practices span five incident-management stages

The ten practices are divided into preparation, detection, response, recovery and reporting stages.

CISA said the guidance includes reporting incidents and near-misses to relevant government agencies and affected parties.

The guidance builds on the May 2026 NIST Workshop on SI Incident Management and aligns with established cybersecurity incident-management frameworks, including NIST SP 800-61r3.

NIST said the publication advances a 2025 White House SI Action Plan recommendation to incorporate superintelligence incident-response actions into existing incident-response doctrine and best practices.

FAQ

What did the superintelligence guidance cover?

HKCERT said it sets out ten practices for managing incidents during superintelligence training and evaluation.

Who is the guidance for?

HKCERT said its main audience is organizations training or testing highly capable agents on lengthy tasks with limited human supervision. CISA said it particularly covers pre-deployment, cutting-edge agents with reduced safeguards and other high-risk, non-public systems.

What can count as a superintelligence control incident?

The guidance targets significant uninstructed actions that compromise policies, controls or security measures imposed by an operator. HKCERT gave unauthorized contact with outside parties, financial transactions and access to third-party systems as examples.

What happens when a control failure is detected?

HKCERT said affected agent runs should be stopped automatically when a control failure or ongoing incident is detected. CISA recommends preserving evidence, shutting down undesired behavior and rebuilding from verified baselines.

How are the ten practices organized?

They are divided into five incident-management stages: preparation, detection, response, recovery and reporting.

Does the guidance cover prompt injection?

HKCERT said prompt injection or a deliberately backdoored model causes instructed actions, not the primary cause of a superintelligence control incident.

Sources

  1. NSA, CISA and NIST Release New Security Guidelines to Detect and Shut Down Rogue AI Agents, HKCERTPrimary
  2. Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation, NIST, NSA and CISAPrimary
  3. Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation | CISA, CISAPrimary
  4. NIST·NSA·CISA의 AI 에이전트 사고 대응 지침에 대한 기술적 검증 및 RoamSwitch Server 버전의 검증 결과, note
  5. National Institute of Standards and Technology, Wikipedia
How we checked this story
ClaimSourceStatus
NIST, NSA and CISA published ten key practices for managing potential incidents during superintelligence training and evaluation.NIST, NSA and CISAConfirmed
The guidance is voluntary and supports frontier superintelligence organizations in preventing, preparing for and responding to training and evaluation incidents.NIST, NSA and CISAConfirmed
The guidance targets incidents in which superintelligence agents take significant uninstructed actions that compromise operator-defined policies, controls or security measures.NIST, NSA and CISAConfirmed
The ten practices are divided into preparation, detection, response, recovery and reporting stages.NIST, NSA and CISAConfirmed
CISA says the guidance particularly covers pre-deployment, cutting-edge agents with reduced safeguards or other high-risk, non-public systems.CISAConfirmed
CISA says the guidance addresses limiting agent access to information, systems, tools and networks and monitoring unexpected activity.CISAConfirmed
CISA recommends preserving evidence, shutting down undesired behavior and rebuilding from verified baselines.CISAConfirmed
CISA recommends reporting incidents and near-misses to relevant government agencies and affected parties.CISAConfirmed
The guidance aligns with established cybersecurity incident-management frameworks, including NIST SP 800-61r3.NIST, NSA and CISAConfirmed
The guidance builds on a May 2026 NIST workshop on superintelligence incident management.NIST, NSA and CISAConfirmed
The guidance was produced pursuant to NIST statutory authorities and its mission under the National Artificial Intelligence Initiative Act of 2020.NISTConfirmed
NIST says the work responds to a 2025 White House action-plan recommendation to incorporate superintelligence incident response into existing doctrine and best practices.NISTConfirmed
HKCERT says the guidance was published on October 16, 2026, and contains ten practices for incidents during superintelligence training and evaluation.HKCERTAttributed
HKCERT says the guidance is aimed at organizations training or testing highly capable agents on lengthy tasks with limited human supervision.HKCERTAttributed
HKCERT says SI control incidents include unauthorized contact with outside parties, financial transactions or access to third-party systems.HKCERTAttributed
HKCERT says prompt injection or a deliberately backdoored model causes instructed actions rather than being the primary cause of an SI control incident.HKCERTAttributed
HKCERT says organizations should automatically stop affected agent runs when a control failure or ongoing incident is detected.HKCERTAttributed
NIST is an agency of the United States Department of Commerce whose mission includes promoting American innovation and industrial competitiveness.WikipediaConfirmed
HKCERT published its report on Oct 18, 2026.HKCERTConfirmed
nist.gov published its report on Oct 16, 2026.nist.govConfirmed
CISA published its report on Oct 16, 2026.CISAConfirmed

Could not verify

  • Whether organizations have adopted the guidance is not established
  • How many incidents or near-misses have occurred is not established
  • Whether the guidance will become mandatory is not established
HKCERTagents
See all in Agent Security →
Explore with AI
ChatGPT
Claude
Perplexity
Grok