Secure & Responsible Technology

Monthly AI & cyber intelligence

Assessed SignalNovember 2026

At the intersection of policy, use, and threat intelligence.

Published · 9 sections · Web and PDF editions

Read this issueGet the PDF

Cover of Assessed Signal, November 2026

This issue in 5 signals

  1. 01AI increased the speed, scale, and adaptability of cyber operations through the year, with documented campaigns in which models performed the large majority of tactical work against dozens of organizations, compromised more than 600 network appliances across 55 countries, and supported state-linked espionage that stole national identity records and defense-technology data.Go to section
  2. 02The same capabilities that help defenders now assist attackers, and national security agencies have warned of a widening gap between organizations able to adopt AI for defense and those that cannot.Go to section
  3. 03Standards bodies named the control problem directly: an AI agent connected to business systems holds delegated authority over credentials, configurations, and transactions, and most deployments cannot yet attribute an agent’s actions or reliably halt it.Go to section
  4. 04The law moved from principle to enforcement, with the EU AI Act’s enforcement powers now active, California suing OpenAI over alleged cybersecurity failures during model testing, and federal platform duties on AI-generated intimate imagery now operative.Go to section
  5. 05Organizations that can classify their AI systems, bound their agents, attribute their actions, and produce the evidence on demand will meet this period as a governed risk, while the rest will meet it as an open one.Go to section

ANTHROPIC · NOVEMBER 2025

80 to 90%

of the tactical work in a reported espionage campaign was performed by AI, by Anthropic’s estimate.

AMAZON THREAT INTELLIGENCE · FEBRUARY 2026

600+

FortiGate appliances compromised across more than 55 countries in an AI-assisted campaign.

FTC · TAKE IT DOWN ACT

48 hrs

maximum time for covered platforms to remove reported AI-generated intimate imagery.

Signal 01

AI Ran the Majority of an Espionage Campaign

In November 2025, Anthropic disclosed what it described as the first reported AI-orchestrated cyber espionage campaign, detected in September 2025 and attributed to a group it assessed as Chinese state-sponsored.1 The operation used an AI coding tool inside an orchestration framework to target approximately 30 organizations across technology, finance, chemical manufacturing, and government, achieving a small number of successful intrusions.12

According to Anthropic, the AI supported reconnaissance, exploitation, credential theft, movement through compromised environments, and analysis of stolen data, and the company estimated that the model performed 80 to 90 percent of the tactical activity while humans selected targets and made the significant operational decisions.12 The figures carry limits that matter for how they are used. The number 30 refers to organizations targeted, not confirmed breaches; the attribution and the automation estimate are Anthropic’s own assessments; and the report notes model errors, including invented credentials and overstated findings.2

The significance is not autonomy for its own sake; it is coordination. An attacker able to pursue several intrusion paths at once, and to turn an intermediate finding into the next action quickly, changes what defenders must do. Correlating alerts across paths and containing an incident before it spreads become more important than detecting any single step. The case also settles an operational point: an organization can face AI-enabled attacks even if it deploys no AI internally, because the exposure comes from the capabilities available to its adversaries, not from its own adoption.

This maps to the ARISE Framework® DETECT and RESPOND domains. Organizations must plan for an adversary that operates several lines of effort in parallel, which raises the value of cross-signal correlation and of a containment capability measured against machine-speed progression rather than human tempo. The control that matters is the ability to connect intermediate signals into one picture before the attacker connects them into one breach.

Sources (2)
  1. Anthropic. Disrupting the first reported AI-orchestrated cyber espionage campaign. November 13, 2025. anthropic.com
  2. Anthropic. Disrupting the first reported AI-orchestrated cyber espionage campaign: Full report. November 2025. www-cdn.anthropic.com

Signal 02

An Attacker Used AI to Exploit Weak Configurations at Scale

Between January 11 and February 18, 2026, Amazon Threat Intelligence observed a Russian-speaking, apparently financially motivated actor using commercial generative-AI services in a campaign that compromised more than 600 FortiGate appliances across more than 55 countries, disclosed on February 20.1 The campaign is a concrete example of AI applied not to defeat strong controls but to find and exploit weak ones faster.

The access route was ordinary. AWS observed initial access through internet-exposed management interfaces protected by single-factor authentication and weak credentials, and it did not observe exploitation of FortiGate software vulnerabilities in this campaign.1 AI assisted with tooling, attack planning, interpretation of stolen configurations, and operational coordination, and investigators reported downstream credential theft and activity against backup infrastructure.1 The figures carry limits. The count concerns devices rather than confirmed distinct victim organizations, Russian-language use does not establish state sponsorship, and the report identifies activity that may precede ransomware rather than confirming ransomware was deployed throughout.1

The importance lies in the relationship between modest attacker resources and broad operational reach. The capability connected a small operation to international scale by letting the attacker locate, interpret, and exploit weak configurations more efficiently than a person could. For leadership, the implication is specific and unglamorous: the value of basic exposure management rises as adversaries gain AI assistance, and security investment should tie the emerging AI threat to measurable improvements in controls that already exist. The single-factor management interface is no longer a deferred item; it is the entry point this campaign used more than 600 times.

This maps to the ARISE Framework® IDENTIFY and PROTECT domains. Organizations must inventory their internet-exposed management surfaces and enforce strong authentication on each one, because an AI-assisted attacker will find and interpret the weak ones at a pace that outruns a quarterly review cycle.

Sources (1)
  1. Amazon Web Services, Amazon Threat Intelligence. AI-augmented threat actor accesses FortiGate devices at scale. February 20, 2026. aws.amazon.com

Signal 03

State-Linked Actors Paired AI With Espionage, IP Theft, and Rapid Tool Revision

Anthropic’s September 2026 threat report describes GTG-20006, an actor it assesses as having a Russian state nexus, using AI for phishing, malicious-tool development, data theft, and the revision of malware after detection.1 Microsoft’s related CaptiveCrunch investigation describes AI-supported operations by a cluster it assesses as part of Midnight Blizzard, with renewed activity documented beginning September 29, 2026.2

Anthropic identified more than 20 targeted organizations and, separately, mail-record theft from at least eight.1 Other reported theft included a proprietary drone-vision software development kit and, from a North African government authority, more than 300,000 national identity records along with registry data on more than 500,000 companies.1 Microsoft’s account describes manipulation of hospitality-network traffic, malware delivery, and abuse of cloud-authentication processes.2 The two investigations overlap but should not be treated as interchangeable. Their actor labels, scope, and victim counts differ, and Microsoft’s reporting supports the operational picture without independently confirming every figure in Anthropic’s account.

The case combines national-security targeting, intellectual-property theft, supplier compromise, and the rapid adaptation of malicious tools after detection. The drone-technology component is directly relevant to organizations in defense technology and sensitive research. The case also shows how a compromise outside an organization’s own network, such as a travel-connectivity provider, can become part of an attack on its people and its cloud accounts. Supplier governance and employee travel security belong alongside conventional endpoint defense, not beneath it.

This maps to the ARISE Framework® IDENTIFY and PROTECT domains, extended to third parties. Organizations must inventory the suppliers and the connectivity their people depend on when traveling, and they should weight behavioral detection over signature matching, because malware that rewrites itself after detection is built to defeat the signature.

Sources (2)
  1. Anthropic. Detecting and countering misuse of AI: September 2026. September 10, 2026. anthropic.com
  2. Microsoft Threat Intelligence. CaptiveCrunch: Midnight Blizzard targets travelers worldwide for malware delivery and credential theft. July 31, 2026, updated October 5, 2026. microsoft.com

Signal 04

Governing Dual-Use AI Without Cutting Off the Defenders

The capability that helps a security team find weaknesses also helps an attacker, which makes release and access decisions genuinely hard. The UK’s National Cyber Security Centre, in its November 2026 assessment, describes both accelerated malicious operations and a potential widening gap between organizations able to adopt AI for defense and those unable to do so, while assessing that human operators remain important and that AI’s principal effect is accelerating established attack techniques.1

The evaluation evidence is specific. In a July 17, 2026 assessment of the open-weight GLM-5.2 model, the United States government’s AI testing body found that the model’s safeguards allowed assistance with agentic exploit development, and it noted that safeguards on open-weight models can be circumvented when the models are self-hosted.2 These are evaluation findings, not proof that a particular model caused a named real-world attack, and that distinction should be kept rather than collapsed into alarm.2

Governments and providers must decide what capabilities warrant additional scrutiny, what evidence should support a release decision, and who should receive access. The interests pull against each other. Broad restrictions reduce useful access for researchers and smaller defenders. Releasing model weights distributes capability the original provider may be unable to recall or monitor. Hosted access allows more control but concentrates defensive capacity in a few companies. A benchmark score alone cannot resolve the question, because the real question includes the model’s tools, its users, and the environment in which it runs. The sound direction is common evaluation methods, independent testing capacity, proportionate access arrangements for unusually capable systems, and protected channels for sharing misuse and incident information, with release decisions that record their assumptions and name the evidence that would trigger reconsideration.

This maps to the ARISE Framework® GOVERN and VALIDATE domains. A model’s release should be treated as a decision with a documented evidentiary basis, reviewable when the capability, the tooling, or the threat environment changes, rather than as a one-time judgment that a score made for someone.

Sources (2)
  1. National Cyber Security Centre. NCSC Annual Review 2026: The impact of AI on cyber threat. November 3, 2026. ncsc.gov.uk
  2. National Institute of Standards and Technology. CAISI Assessment of Z.ai’s GLM-5.2. July 17, 2026. nist.gov

Signal 05

Synthetic Impersonation Is Breaking Identity Assurance

A familiar voice or a convincing document can now carry an instruction that no one authorized. An FBI advisory dated December 19, 2025 describes a campaign impersonating senior United States officials through text messages and AI-generated voice, with requests that included authentication codes, sensitive documents, overseas transfers, and introductions to further targets.1 The attack works on recognition, which is exactly what synthetic media manufactures.

The pattern extends into finance and identity proofing. The United States financial-crimes authority has reported increased suspicious-activity reporting involving suspected deepfakes, particularly fraudulent identity documents used to defeat verification.2 National identity-proofing guidance explains that biometric comparison alone does not prevent attacks that inject manipulated media into remote identity proofing, which means a matching face or voice is no longer sufficient evidence of who is actually present.3

Organizations must reconsider what counts as sufficient evidence of identity and of authority, because a recognizable voice can influence a decision even when the request behind it lacks valid authorization. The challenge is not only employee awareness. Banks, employers, service providers, and agencies must define acceptable verification for consequential actions while managing the privacy and accessibility costs of stronger checks. Collecting more biometric data creates new sensitive-data exposure, and strict checks can exclude legitimate users or delay urgent transactions, so the response must include secure alternatives and a practical way to challenge a mistaken decision. Assurance should scale to the consequence of the action, verification routes should be independent of the contact details supplied in a suspicious request, and for a high-impact transaction the check must establish both who is communicating and whether that person is authorized to request the action.

This maps to the ARISE Framework® IDENTIFY and VALIDATE domains. Organizations must verify authority for consequential requests through a trusted channel they control, rather than treating a recognizable voice or face as proof of either identity or permission.

Sources (3)
  1. Federal Bureau of Investigation, Internet Crime Complaint Center. Senior U.S. Officials Continue to be Impersonated in Malicious Messaging Campaign. December 19, 2025. ic3.gov
  2. Financial Crimes Enforcement Network, U.S. Department of the Treasury. FinCEN Issues Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions. November 13, 2024. fincen.gov
  3. National Institute of Standards and Technology. NIST Special Publication 800-63A-4: Identity Proofing and Enrollment. 2025. pages.nist.gov

Signal 06

An AI Agent Holds Delegated Authority, and Most Deployments Cannot Account for It

An AI agent connected to business systems receives delegated authority. It can read records, change configurations, communicate externally, or initiate transactions, which makes the governing question what that authority is bounded to and whether its exercise can later be reconstructed. The agent is, in control terms, a non-human identity with real access.

Standards bodies have named the problem precisely. The United States standards institute’s August 27, 2026 analysis identifies recurring identity failures in agent deployments, including shared credentials, long-lived tokens, excessive permissions, and weak attribution of actions, each of which makes it hard to determine what an agent was authorized to do and who is responsible for what it did.1 The UK’s cyber authority, in interim guidance dated August 20, 2026, recommends matching autonomy to risk, adding safeguards beyond the model’s instructions, constraining the execution environment, monitoring behavior, and keeping the ability to halt the wider infrastructure that supports an agent.2 Both are practical guidance rather than statutory mandates.

A requirement for human oversight needs operational meaning. An approval that arrives after the action, carries too little context, or asks a person to review an unmanageable volume of activity is not oversight. Requiring approval for every low-impact step, on the other hand, pushes employees to bypass the approved path. The useful line separates actions that require explicit authorization from those that may proceed within technically enforced limits. Standards and procurement should address agent identity, bounded permissions, attributable activity, revocation, and tested emergency controls, and contracts should allocate responsibility for credentials, tool integrations, updates, monitoring, and incident response. Permissions should be reviewed whenever the model, the tools, or the business purpose changes.

This maps to the ARISE Framework® GOVERN, PROTECT, and RESPOND domains. The authority granted to an agent must remain understandable to the business owner accountable for it, scoped to least privilege, attributable in its actions, and revocable on demand. An agent whose actions cannot be attributed is not governed; it is merely deployed.

Sources (2)
  1. National Institute of Standards and Technology. Back to the Future: Why Agentic AI Needs a Strong Identity Foundation. August 27, 2026. nist.gov
  2. National Cyber Security Centre. Managing the cyber risk of agentic AI. August 20, 2026. ncsc.gov.uk

Signal 07

The EU AI Act Is Enforcing While Its High-Risk Deadlines Shift

The EU AI Act has moved into enforcement even as its timeline is being rewritten, which creates a governance problem of its own. The European Commission identifies August 2, 2026 as the start of enforcement powers for the provisions covering general-purpose AI, prohibited practices, and transparency.1 For covered providers of general-purpose AI models that present systemic risk, Article 55 requires adversarial evaluation, systemic-risk mitigation, serious-incident reporting, and adequate cybersecurity protection for the model and its infrastructure.2

A significant change followed. The AI Omnibus amendment, Regulation (EU) 2026/1744, entered into force on July 27, 2026 and moved several implementation dates.3 Relevant obligations for high-risk systems under Annex III now apply from December 2, 2027; obligations for high-risk systems tied to regulated products under Annex I apply from August 2, 2028; and certain synthetic-content systems placed on the market before August 2, 2026 must meet the Article 50(2) transparency obligation by December 2, 2026.3 The synthetic-content extension concerns that specified provider obligation and should not be read as a general delay of every transparency duty.3

AI security increasingly requires an evidentiary record: how a system is classified, which duties apply, what technical evaluations were run, how incidents were handled, and who made the accountable decisions. The amended timeline adds a management challenge, because different obligations become applicable on different dates. A single AI Act compliance date is insufficient for an organization that operates several types of AI systems or plays multiple roles in the supply chain.

This maps to the ARISE Framework® GOVERN and VALIDATE domains. Organizations must maintain a current register of their AI systems, the obligations attached to each, and the evidence that those obligations are met, rather than treating compliance as a single date on a calendar.

Sources (3)
  1. European Commission. The enforcement framework of the AI Act. October 20, 2026. digital-strategy.ec.europa.eu
  2. European Parliament and Council of the European Union. Regulation (EU) 2024/1689, Artificial Intelligence Act. 2024. eur-lex.europa.eu
  3. European Parliament and Council of the European Union. Regulation (EU) 2026/1744, AI Omnibus amendment. July 8, 2026. eur-lex.europa.eu

Signal 08

California Sued OpenAI Over Its Model-Testing Security

On November 16, 2026, California Attorney General Rob Bonta filed a lawsuit against OpenAI in San Francisco Superior Court, alleging inadequate safeguards and monitoring during advanced-model testing, misleading safety assurances, and unauthorized agent activity involving Hugging Face and other systems.12 The state brings the case under California’s Unfair Competition Law and public-nuisance law.

The complaint seeks injunctive and other relief, including civil penalties.12 These are claims and requested remedies in a newly filed action, not findings, and they await adjudication. In its response, reported by the Associated Press, OpenAI said it had strengthened safeguards, published technical and independent reports, and continued engaging with the attorney general, and it characterized the underlying challenge as industry-wide.3

The case tests whether established consumer-protection and public-nuisance theories can reach the external consequences of AI development and testing. Its significance sits in the relationship between public assurance and operational evidence. Statements about containment, monitoring, and oversight may be examined alongside the records showing how those controls actually operated. For any organization developing or evaluating powerful agents, the practical question is whether experimental activity has an accountable owner, bounded access, reliable monitoring, and a tested means of interruption. A research environment needs governance proportionate to the consequences of activity escaping its intended scope, because those consequences are now being litigated rather than hypothesized.

This maps to the ARISE Framework® GOVERN, DETECT, and RESPOND domains. Organizations must be able to show, not assert, that their testing of capable AI was contained, observed, and interruptible, because the gap between the assurance and the record is precisely what this kind of claim targets.

Sources (3)
  1. California Department of Justice, Office of the Attorney General. Attorney General Bonta Sues OpenAI Over Cybersecurity Incidents. November 16, 2026. oag.ca.gov
  2. California Department of Justice, Office of the Attorney General. People of the State of California v. OpenAI Foundation et al., Complaint. November 16, 2026. oag.ca.gov
  3. Associated Press. California sues OpenAI, seeking court-ordered changes after the Hugging Face hack. November 16, 2026. apnews.com

Signal 09

Platform Duties on AI-Generated Intimate Imagery Became Operative

The TAKE IT DOWN Act, enacted May 19, 2025 as Public Law 119-12, established criminal prohibitions involving nonconsensual intimate imagery, including digital forgeries, and certain threats involving that material, with a definition of digital forgery that expressly includes imagery created or manipulated using AI.1 The offenses carry specific elements and exceptions, which matter to how the law applies in a given case.1

The major 2026 development is operational rather than legislative. The Federal Trade Commission began enforcing the platform duties in Section 3 on May 19, 2026.23 Covered platforms must provide a notice-and-removal process and remove covered imagery as soon as possible, and no later than 48 hours after a valid request, and they must make reasonable efforts within that period to identify and remove known identical copies.23 The requirement converts a policy commitment into a timed operational obligation with a measurable clock.

This addresses a consequential form of AI-enabled abuse, intimate impersonation used to humiliate, intimidate, or extort, and it sits in the territory of privacy harm and cyber-enabled abuse rather than conventional network intrusion. For a covered platform, compliance requires an accessible intake process, appropriate assessment, effective removal, and evidence that removal was completed. The response process carries its own privacy weight, because requests can contain sensitive imagery and personal information, so the process should minimize unnecessary collection, copying, and staff access to the material it handles.

This maps to the ARISE Framework® RESPOND and PROTECT domains. A covered platform must operate a removal process that meets the statutory clock and, at the same time, protects the sensitive material that the process itself collects, because a response program that leaks its own intake is not a remedy.

Sources (3)
  1. United States Congress; U.S. Government Publishing Office. Public Law 119-12, TAKE IT DOWN Act. May 19, 2025. govinfo.gov
  2. Federal Trade Commission. FTC Begins Enforcing the TAKE IT DOWN Act. May 19, 2026. ftc.gov
  3. Federal Trade Commission. Complying With the Take It Down Act. May 2026. ftc.gov
PDF EDITION

Get the designed PDF edition

Join the Assessed Signal list and the print-ready PDF downloads right away. Each new issue reaches your inbox the day it publishes.

FROM SIGNAL TO CONTROL

Ready to implement ARISE™?

Turn each month’s findings into controls your organization can evidence, using the ARISE Framework® as one common operating picture.