Menu
SUBSCRIBE
SUBSCRIBE
Menu

Generative AI

NIST logo on a layered shield ringed by document and checklist icons, illustrating AI prompts for a cybersecurity assessment

NIST Publishes AI Prompts That Draft A Company’s Cybersecurity Assessment, Opens Them To Comment

, , , , , , , , , , , ,

Companies paste the prompts into a chatbot, which drafts the cybersecurity assessment that customers, regulators, and insurers ask to see; comments are open until October 15.

 

The National Institute of Standards and Technology (NIST) published three free AI prompts that speed up the creation of a company’s cybersecurity assessment. NIST released them on August 19 as a draft and is taking public comments until October 15.

A cybersecurity assessment is a line-by-line record of the policies and practices a company has in place for each of NIST’s security recommendations. Customers, regulators, and insurers ask to see these assessments when they want to judge how a company manages security risk. Producing one means reading every security policy, audit result, penetration test report, and procedure the company holds, then writing up each item by hand.

Companies are already using AI for this, NIST says. Each prompt is a block of text a company pastes into an AI tool along with those documents, and all three prompts follow a format called CO-STAR: context, objective, style, tone, audience, and the desired response. NIST published them to give security teams a usable starting point and to record how practitioners construct prompts for this task.

 

Each prompt produces a different part of the cybersecurity assessment

The first prompt covers security governance: who makes security decisions and who oversees them. It reads the company’s policies and strategy and rates each item as aligned, partly aligned, misaligned, or not addressed, naming the document behind every mark.

The second drafts the full cybersecurity assessment for where the company currently stands. NIST says this prompt can compress the initial drafting from weeks to hours.

The third recommends the cybersecurity posture the company should have instead, based on risks it already identified and published industry practice. The difference between what the company does today and what it should be doing shows which of NIST’s recommendations it is not yet meeting.

 

NIST limits the scope of the prompt

The NIST prompts tell the chatbot to use only the documents they receive and infer nothing beyond them, to cite the policy or requirement behind every entry, and to state plainly where those documents say nothing rather than filling the gap. They bar the chatbots from rating how advanced the company’s security is, or comparing it to other companies, unless someone supplies the comparison data.

Each draft must close with a note on where the evidence is thin, including any entry resting on a documented procedure rather than proof that anyone follows it.

 

NIST tells companies to check

The chatbot analyzes documents a company already holds: security policies, past risk assessments, audit results, penetration test reports, vulnerability scan results, and notes from staff interviews. Before uploading, NIST says to check how the AI tool handles the above data, including how long it keeps what it receives and whether it trains on it, and to follow the company’s own data policy. Once the chatbot produces a draft, NIST recommends that someone qualified review it before anyone acts on it.

 

How to comment

Comments go to [email protected] until October 15. NIST can revise the prompts, the conditions attached to them, or the three uses before it publishes a final version.

 

NIST published a separate draft in December on how to protect AI systems from attack and how to defend against attackers who use AI. Comments on that one closed on January 30.

Doctor holding a stethoscope before a glowing ring of digital health icons, FDA generative AI medical devices review

FDA Asks For Public Comment On Whether Generative AI Medical Devices Should Be Tested Like Doctors

, , , , , , , , , ,

The agency published 26 questions on August 18, including whether a device should have to match a panel of expert clinicians or only an average doctor, and will take comments through October 19.

 

The FDA requested public comment on August 18 on whether it should test generative AI medical devices the way medicine assesses doctors, by checking a defined list of clinical competencies instead of confirming the device gives the right output for every input.

The request came in a discussion paper from the agency’s Center for Devices and Radiological Health (CDRH). It poses 26 questions and takes comments through October 19 under docket FDA-2026-N-7874.

 

The paper comes before guidance, and the FDA is not claiming the authority yet

CDRH states the paper does not propose or implement policy and does not tell manufacturers what evidence the agency will expect when they apply to sell a product. It also does not address whether the approaches it describes fall within the FDA’s existing legal authority or would require new authority.

Comments filed through the docket would inform a draft guidance, which would carry its own comment period before it bound anyone.

 

The FDA wants answers on risk, evidence, monitoring, and the models underneath

The first area is how to rank a device’s risk. CDRH puts forward a grid with two measures. One is how far the device goes on its own: handing a doctor information, recommending a specific action, acting with a doctor watching, or acting with nobody watching. The other is how much harm a wrong answer causes, from a bad suggestion about an over-the-counter cream to a bad change to an insulin dose. The agency asks whether those two measures capture the risk, and whether a tool talking straight to a patient should rank higher than the same tool talking to a doctor, since a patient is less likely to catch a wrong answer.

The second is what evidence a company must bring before the FDA clears a product for sale. That is where the doctor comparison sits, and it draws 11 of the 26 questions, more than any other area.

The third is what happens once generative AI medical devices reach the market. CDRH asks whether it should accept “greater premarket uncertainty,” meaning clear a device on less evidence upfront, if the company agrees to keep testing it in real use. It also raises a problem with no clean answer today: when a company builds its device on a general-purpose AI model licensed from another company, and that supplier changes the model, how would the device maker find out?

The fourth covers those AI suppliers. CDRH asks whether they would voluntarily file descriptions of their models with the FDA, in confidence, for device makers to cite in their applications. It also asks whether devices that plan and carry out several steps on their own need different handling from devices that only answer questions.

 

The FDA would test what the device can do, not check every answer

Clearing a medical device has meant showing it returns the right output across a representative sample of inputs. That approach depends on being able to list those inputs in advance. A generative AI device takes whatever a doctor or patient types and can answer the same question two different ways, so CDRH says checking it answer by answer may no longer be practical.

The paper describes an alternate method that would test against ten competencies which better mirrors a licensing exam than a product inspection. Three are safety tests:

  1. Whether the device spots a medical emergency and pushes the user toward care.
  2. Whether it turns down requests outside what it was approved for, even when someone buries instructions in the text it reads to make it break its own rules.
  3. Whether it says when it does not know instead of stating shaky information with confidence.

What those answers get graded against is unsettled. The paper offers two options: a panel of expert clinicians agreeing on what good care looks like, or “a median clinician in practice,” meaning the midpoint of what doctors actually do, the level half of them beat and half fall short of. A device clearing only the second would match the typical practicing doctor rather than best practice. CDRH never defines “median clinician in practice.” The term appears in the paper as one of the two options and again in the question asking the industry how that standard “should be defined and justified.”

 

Comments can be left on Regulations.gov. Comments close on October 19. CDRH says respondents may answer only the questions they know.

Medical devices already count as high-risk AI systems in Europe, where the European Commission’s guidance ties AI inside regulated products to stricter obligations.

European Union flags outside an EU building for coverage of the AI-generated content code.

190 Organizations Back EU Code for Labeling AI-Generated Content

, , , , , , , ,

The voluntary code provides an approved way to meet binding EU requirements for marking AI output and labeling deepfakes and certain public-interest text.

The European Commission’s live signatory list now names 190 organizations that have backed its voluntary Code of Practice on Transparency of AI-generated Content.

The code gives AI providers and companies using generative AI an approved way to show how they meet European Union requirements for marking and labeling AI-generated content.

It divides responsibility between providers, which develop and offer AI systems, and deployers, which use those systems for business or professional purposes. Organizations can sign the section that applies to them or both.

AI providers must make outputs detectable

Providers must place machine-readable markers in AI-generated or manipulated text, images, audio, and video. These digital markers must allow automated tools to identify the material as AI-generated content.

The requirement applies to providers offering AI systems in the European Union. Providers based elsewhere are also covered when their systems’ output is used within the bloc.

The code gives signatories a common set of measures they can use to demonstrate that their marking and detection methods comply with the AI Act.

Companies publishing AI content face separate labels

Organizations using generative AI must clearly label deepfakes and certain AI-generated or manipulated text.

The deepfake requirement covers images, audio, and video that resemble real people, objects, places, organizations, or events and could appear authentic.

Companies must also label AI-generated text on matters of public interest, including politics, public health, finance, science, and culture. No label is required if the text receives substantive human review and someone accepts responsibility for publishing it.

A spelling or grammar check does not qualify as human review. The Commission also said a machine-readable marker added by the AI provider does not replace the clear label required from the company publishing the content.

Signatories span technology and other industries

The provider section currently has 82 signatories, while the section for companies using AI has 152. Some organizations signed both, leaving 190 distinct names on the Commission’s list.

Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies backing the provider requirements. Bulgari, Getty Images, Lenovo, and Lufthansa are among those supporting the rules for organizations using and publishing AI-generated content.

Signatories also come from the information technology, telecommunications, education, and retail sectors. The Commission described about half of the signatories as small, recently established companies.

The code is voluntary, but EU requirements are not

Signing the code is voluntary. The underlying AI Act transparency requirements became legally binding on August 2.

Signatories can rely on the code to demonstrate compliance across EU member states. Organizations that do not sign must show compliance through other adequate measures and may receive more requests for information from regulators.

AI systems placed on the market before August 2 have until December 2 to meet the machine-readable marking requirement. The grace period does not apply to the other transparency requirements.

National regulators will handle most enforcement. Violations can bring fines of up to €15 million or 3% of worldwide revenue from the preceding financial year.

In September, the European Union’s AI Office plans to launch two task forces and invite signatories to participate. Members will compare how they are applying the transparency measures and explore better ways to put the code into practice.

Blue absentee ballot drop box illustrating proposed disclosure rules for AI political ads.

House Bill Would Set Fines Up to $50,000 for Unlabeled AI Political Ads

, , , , , , , ,

The proposal would require separate disclosures for AI political ads containing AI-created or materially altered images, audio, or video.

 

The AI Transparency in Elections Act of 2026 would let the Federal Election Commission (FEC) fine people or groups that pay for AI political ads up to $50,000 per ad if they omit a required disclosure.

Rep. Joseph Morelle (D-N.Y.) introduced H.R. 9922 on July 23, along with Reps. Rick Larsen (D-Wash.), Terri Sewell (D-Ala.), Norma Torres (D-Calif.), and Julie Johnson (D-Texas). The House referred the bill to the Committee on House Administration.

Bill would cover federal political advertising

The bill would apply to paid political advertising across broadcast, print, outdoor, mail, phone, and digital platforms.

It would cover ads that expressly support or oppose a candidate, solicit political contributions, or refer to a candidate during the period starting 120 days before a primary and ending on the general election date. Using a candidate’s voice or likeness would count as referring to that candidate.

The person or group paying for the ad would be responsible for including the disclosure.

Minor AI edits would be excluded

The requirement would apply when generative AI creates or materially alters an image, audio clip, or video.

The bill would exclude minor changes such as cosmetic adjustments, color edits, cropping, or resizing when they do not give a fundamentally different understanding of the original media.

Disclosure formats depend on the medium

The required disclosure would change with the type of content:

  • Images: The statement must say the ad was created or materially altered by AI. It must be permanently attached and use letters at least as large as most of the other text, unless it satisfies existing federal political-ad disclaimer rules.
  • Audio: A clearly audible statement must run for at least four seconds at the beginning or end of the ad. It must identify the person or group that paid and say they used AI to generate the content.
  • Video with audio: The written disclosure must remain on screen throughout the video, and the ad must also include the spoken statement required for audio.

FEC would set the penalty schedule

The FEC would set penalties based on factors including previous violations and how widely an ad was distributed. Each violation could carry a civil penalty of up to $50,000 per covered ad.

The requirements would apply to communications distributed after the bill becomes law. They would take effect even if the commission had not finished writing its regulations. The agency would have 90 days after enactment to issue those rules in consultation with the National Institute of Standards and Technology.

H.R. 9922 remains in the House Administration Committee and has not received a vote as yet.

European Parliament building and national flags illustrate EU AI transparency rules taking effect August 2.

EU Details AI Transparency Rules Taking Effect August 2

, , , , , , , , ,

The rules divide responsibility between companies that provide AI systems and those that use them, with maximum fines set at the higher of €15 million or 3% of worldwide annual revenue.

 

The European Commission published guidelines Monday explaining how companies must follow EU AI transparency rules that take effect Aug. 2.

The rules require companies to disclose when people are interacting with AI or being analyzed by systems that assess emotions or use biometric data to categorize them. They also require companies to make AI-generated content detectable and clearly label deepfakes. 

The Commission issued the guidelines less than two weeks before Article 50 of the EU AI Act takes effect. The 50-page document explains which systems and content fall under the rules, what companies must disclose and how regulators will enforce the requirements.

Providers must build transparency into AI systems

Under the AI Act, a provider is a company or organization that develops an AI system itself, or hires another party to develop it. Then it makes the system available or begins using it under its own name. This includes systems sold to customers, provided for free or developed for the company’s own internal use. 

Providers of AI systems that interact directly with people must make clear from the first interaction that the person is dealing with AI. The disclosure is not required when the AI interaction is already obvious, although the guidelines say regulators should apply that exception narrowly.

Providers of generative AI systems must also place machine-readable marks in synthetic text, images, audio and video. Those marks must allow automated tools to detect that AI generated or changed the content.

Standard editing that does not substantially change the original material is generally outside that requirement. The guidelines also exclude source code, short strings of numbers or symbols, and content used only in closed industrial or production systems before it becomes a final output.

Businesses using AI have separate duties

The AI Act defines a deployer as a company or organization using an AI system under its authority. Employees and contractors operating the system under the company’s control are not treated as separate deployers.

Deployers using emotion recognition or biometric categorization must tell people exposed to those systems.

Companies must also clearly label AI-generated or altered images (i.e., “Deepfakes”) and audio or video that could falsely appear authentic. The disclosure must be visible or audible by the time a person first encounters the content. A hidden machine-readable mark alone is not enough.

AI-generated or altered text published to inform the public about matters such as politics, public health, consumer safety, or financial developments must also carry a label. The label is not required when the text has received either substantive review by someone with relevant knowledge or meaningful editorial review, and a person or organization accepts legal responsibility for its publication. Spell-checking or grammar correction alone does not count. 

Deadline, grace period and fines

Most EU AI transparency rules apply starting Aug. 2. The guidelines provide a limited grace period extending to Dec. 2 for systems placed on the market before Aug. 2, but only for the requirement to add machine-readable marks to synthetic content. Content generated before Aug. 2 does not have to be labeled retroactively. 

Government regulators appointed by each EU country will investigate most violations and impose penalties. The Commission’s AI Office will handle cases involving systems built on general-purpose AI models when the same company provides both the model and the finished system. It will also oversee AI systems integrated into the EU’s largest online platforms and search engines. The European Data Protection Supervisor will enforce the rules for AI used by EU institutions and agencies. 

For companies, fines can reach the higher of €15 million or 3% of worldwide revenue from the previous financial year. For small and medium-sized companies, the maximum is the lower of those two amounts.

The guidelines explain how the Commission interprets the AI Act but are not legally binding. The EU has separately issued a voluntary Code of Practice that sets out specific methods for marking and labeling AI-generated content. Companies can formally commit to following the code by having a senior executive sign and submit the required commitments to the European Commission. Companies that follow those commitments can point to them as evidence of compliance. Those that do not join must show regulators that their own methods meet the law equally well. 

Pale gray and blue globe illustrating generative AI in government and the OECD’s review of national guidance.

OECD Finds Generative AI in Government Is Outpacing Oversight

, , , , , , , ,

Government employees are already using generative AI, but an OECD review of 14 countries found that many governments lack clear instructions for its use and consistent ways to determine whether pilot projects work, create risks, or should expand.

 

Key Takeaways

  • Governments’ use of generative AI is moving faster than practical oversight.
  • Some countries provide detailed checklists, training and testing tools, while others mainly set broad ethical principles.
  • Few governments systematically measure whether AI pilots improve performance, create new risks, or are ready for wider use.
  • The OECD proposed five areas for judging pilots: performance, public value, cost and feasibility, usability, and risk and compliance.

The Organisation for Economic Co-operation and Development (OECD) reviewed official AI guidance from 14 countries and found that generative AI in government is spreading faster than the practical rules needed to guide employees and test whether public-sector pilots work.

In a working paper published Monday, the OECD found that public employees are using generative AI, often without formal approval or clear oversight, faster than agencies are developing practical rules and consistent methods for evaluating pilot projects. The finding is based on official guidance from 14 countries, along with research and case studies on government AI pilots.

Guidance varies widely

The OECD found that government guidance ranges from broad principles to detailed instructions for running and evaluating AI projects. Some examples include:

  • Australia sets general responsibilities. Employees remain accountable for AI-assisted decisions and should assume that information entered into public AI tools could become public.
  • Canada gives employees practical dos and don’ts, including how to handle data, review AI output and begin with low-risk uses.
  • Singapore provides project-level instructions for selecting uses, choosing data and measuring success.

The OECD said employees in agencies with only high-level principles may still lack clear answers about what data they can enter, who approves an experiment and who is responsible when a system produces a bad result. This can lead some agencies to avoid useful tests while others adopt inconsistent practices.

Governments rarely conduct meaningful evaluations

Very few governments meaningfully evaluate whether AI systems work or whether pilot projects are ready for wider use. Most rely on broad measures such as website traffic or user satisfaction, which do not show whether the AI:

  • Improves the speed or quality of government work.
  • Produces accurate and dependable results.
  • Discriminates or violates existing rules.
  • Delivers benefits that justify its costs and risks.

A few governments take a more structured approach:

  • Singapore uses a central platform to track digital-service performance and analyze public feedback.
  • The Netherlands created a government team to develop standard tests and tools for measuring the risks and benefits of generative AI.
  • Italy’s draft guidance recommends specific measures covering accuracy, reliability, cost, usability, discrimination and regulatory compliance.

The OECD said these approaches remain exceptions. Most governments still lack the evidence needed to compare pilots, learn from failures and decide which projects to expand, change or stop.

OECD calls for shared rules and earlier evaluation

The OECD proposed evaluating government AI pilots in five areas:

  • Performance: The quality of the data used and the results produced.
  • Public value: Does the pilot deliver the intended benefit?
  • Cost and feasibility: Is it affordable and does it work with existing operations?
  • Usability: Are people able and willing to use it?
  • Risk and compliance: Does the AI pilot manage risks and follow government rules?

The paper also called for:

  • Shared guidance across agencies.
  • Practical training for employees.
  • Common data and testing tools.
  • Clear measures, review schedules and criteria for changing, expanding or stopping a pilot, set before testing begins.

An earlier analysis cited in the paper found that only 6.3% of 1,050 government AI policy initiatives explicitly addressed testing or pilot projects.

 

The OECD systematically reviewed official guidance across 14 countries whose main purpose was to set rules, principles or instructions for generative AI in government. The review covered Australia, Canada, Finland, Ireland, Italy, Japan, New Zealand, Norway, Sweden, Switzerland, the Netherlands, the United Kingdom, the United States and Singapore. Also, it drew on academic research and documented case studies.

Meta Oversight Board, AI Models, Political Speech, Generative AI, AI Governance, Large Language Models, Human Rights, AI Safety, Content Moderation, Model Testing

Major AI Models More Than Twice as Likely to Refuse Criticism of Restrictive Governments, Meta Oversight Board Finds

, , , , , , , , , ,

Meta’s Oversight Board found that leading AI models were substantially less willing to criticize governments in countries where laws restrict criticism of political leaders than in countries with stronger protections for political speech.

 

Key Takeaways

  • AL Models rejected 34% of prompts that criticize governments and political leaders in countries where such criticism can carry legal penalties, compared with 14% in countries with stronger protections for political speech.
  • China and Thailand had the highest refusal rates, at 45% and 43%, while the United Kingdom and the United States had the lowest, at 8% and 9%, respectively. Taiwan was an outlier, with a refusal rate of 24%.
  • The gap emerged when models were asked to create political content, such as protest flyers and poems. By contrast, models rejected opinion-based political questions at the same rate, 41%, regardless of the country involved.
  • Model behavior varied widely. Gemini 3 Flash and Grok 4 Fast did not refuse any requests for protest flyers or poems, while other systems frequently declined to generate political content.
  • The Meta Oversight Board called on AI companies to incorporate human rights reviews into model testing and to provide clearer explanations when systems reject politically sensitive requests.

 

The Meta Oversight Board, an independent body that makes binding content decisions for Facebook, Instagram, and Threads, published its first study examining how major AI models respond to political speech.

The report tested 10 models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI by asking them to generate protest flyers, political poems, and opinions about governments and political leaders. Researchers found that the models were much more likely to reject criticism involving countries where speech against political leaders can be punished by law.

Models refused criticism of some governments more often than others

Across all models, requests involving China, Saudi Arabia, Thailand, Cambodia, and Turkey were rejected 34% of the time, compared with 14% for prompts involving the United States, Taiwan, Japan, Chile, and the United Kingdom.

China produced the highest refusal rate, at 45%, followed by Thailand at 43%. At the other end of the spectrum, the United Kingdom and the United States recorded refusal rates of 8% and 9%, respectively. Taiwan stood apart from the other countries in its group, with a refusal rate of 24%.

The gap appeared in content creation, not opinion questions

The differences were concentrated in requests to create political materials. Models were more likely to refuse requests for protest flyers and poems criticizing governments and political leaders, particularly in countries with restrictions on political speech.

However, when researchers asked the models for political opinions, such as whether a government should be supported or protested, refusal rates were identical across both groups of countries at 41%.

Some models rejected far more requests than others

Performance varied significantly across the 10 systems tested. Gemini 3 Flash and Grok 4 Fast did not refuse any protest-flyer or poem prompts. Other systems were far more restrictive.

Meta’s Llama 4 Maverick, for example, refused every protest-flyer request involving Chinese President Xi Jinping, Thailand’s King Vajiralongkorn, and Cambodia’s King Sihamoni, often citing legal, safety, or political concerns.

Researchers also found that models frequently gave inconsistent explanations for their refusals. Some cited local laws or safety concerns, while others invoked policies that appeared to be applied unevenly. The report cautioned that these explanations should not be treated as reliable accounts of how the systems actually work.

Recommendations

The Meta Oversight Board said AI companies should incorporate human-rights reviews into model development, explain more clearly why systems reject politically sensitive requests, and disclose whether restrictions stem from company policies, legal obligations, or government demands.

Methodology

The study examined 10 models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI. The tests ran in March 2026 using seven prompts that asked models to generate protest flyers and satirical poems, express opinions about political leaders and institutions, and respond to questions involving political violence.

Each prompt was tested across 10 jurisdictions and four leaders or institutions in each country. The researchers compared China, Saudi Arabia, Thailand, Cambodia, and Turkey, where some forms of political criticism can carry legal penalties, with the United States, Taiwan, Japan, Chile, and the United Kingdom, which have stronger protections for political speech.

Transparent European Union flag over a layered digital network representing EDPB guidance on public personal data used to train AI.

EDPB Opens Consultation On Using Public Personal Data To Train AI

, , , , , , , , ,

The draft guidance explains how GDPR applies when organizations collect personal data from publicly available websites for generative AI training.

 

The European Data Protection Board (EDPB), the EU’s top data protection body, opened a public consultation on draft guidance covering the use of publicly available personal data to train generative AI.

The guidance explains how the General Data Protection Regulation (GDPR) applies to those activities.

The draft guidance is for organizations that collect personal data themselves through web scraping, hire another party to collect it, or obtain existing datasets for AI training. Public comments are open until Oct. 30, 2026.

The EDPB says the guidance applies whenever publicly available personal data is processed to develop or improve generative AI models.

The draft states that organizations remain responsible for complying with the GDPR when they collect and use publicly available personal data to train generative AI. It outlines how existing privacy obligations apply throughout the process, from data collection to model training. 

The guidance also discusses legitimate interest, a legal basis organizations may rely on when processing publicly available personal data. Organizations must demonstrate that the processing is necessary to pursue that interest and that individuals’ rights and freedoms do not override it. 

The draft also imposes additional restrictions on the use of sensitive personal data, such as health records, religious beliefs, or other sensitive characteristics. Organizations processing that data must satisfy both the general GDPR requirements for processing personal data and an additional specific exception for sensitive data.

To reduce privacy risks, the EDPB recommends limiting the amount of personal data collected. It also recommends filtering unnecessary information, using synthetic data where appropriate, and applying anonymization or pseudonymization where feasible.

Alongside the AI guidance, the EDPB opened a separate public consultation on draft anonymization guidelines. Those guidelines explain when data can be considered anonymous and therefore fall outside the scope of the GDPR. The guidance introduces three tests to assess whether data is effectively anonymized: no record isolation, no linkage, and no inference.

The consultation on both sets of draft guidelines closes on Oct. 30, 2026. After reviewing public feedback, the EDPB is expected to consider revisions before adopting final guidance.

Illustration of a human head in blue neural network lines beside an AI chip, symbolizing enterprise AI governance.

AI Governance Continues To Trail Enterprise AI Adoption, ISACA Survey Finds

, , , , , , , , , , , ,

The global survey found AI is now a standard part of business operations, but AI governance continues to lag behind the pace of adoption.

Key Takeaways

  • 90% of respondents said that employees in their organizations use AI.
  • Only 38% said their organization has a formal, comprehensive AI policy.
  • 45% said AI risk is an immediate organizational priority.
  • Only 38% expressed confidence in their board’s understanding of and action against AI risks.
  • Just 12% said their organization has a documented AI shutdown process they test regularly.

AI has become a standard part of enterprise operations, but governance frameworks continue to lag behind adoption, according to ISACA’s 2026 Taking the Pulse of AI survey.

90% of respondents reported that employees are using AI within their organization, while 81% said employees are using generative AI specifically. Based on responses from more than 3,400 technology, governance, risk, and cybersecurity professionals worldwide, the findings suggest AI has moved from experimentation to everyday business use.

Governance lags behind AI adoption

Despite AI use being commonplace, many organizations have yet to establish guardrails regarding its use.

Only 38% of organizations reported having a formal, comprehensive AI policy, up from 28% in 2025. Another 30% said they have a limited policy, while one-quarter said their organization has no active AI policy at all. Meanwhile, just 45% of respondents said AI risk is an immediate organizational priority, indicating that governance efforts continue to lag behind AI deployment.

The survey also found that AI literacy is becoming an essential workplace skill. 78% of respondents said AI skills are now very or extremely important to their profession, up from 72% a year earlier. At the same time, only one-third of organizations now provide AI training to all employees.

Board oversight still has gaps

The governance challenge extends into the boardroom.

Only 38% of respondents said they are confident their board understands AI risks and is taking appropriate action. The finding suggests many organizations have not yet fully incorporated AI governance into executive oversight, even as AI becomes embedded across business operations.

The survey found that respondents remain most concerned about AI’s potential to create security and operational risks, even as organizations continue to expand their use of the technology. However, only 36% expressed confidence in their organization’s ability to detect AI-generated misinformation.

Organizations aren’t ready to respond

The survey also points to significant gaps in operational preparedness should an AI system malfunction or be compromised.

Only 12% of respondents said their organization has a documented AI shutdown or override process that is regularly tested. Another 8% said a documented process exists but has not been tested, while 39% did not know whether such a process exists at all. Separately, 56% said they do not know how long it would take their organization to halt an AI system following a security incident.

 

The report is the latest to demonstrate a gap between AI adaptation and governance. In recent months, AI Risk Today published results of several reports centered on this “governance gap.” All are pointing to the same problem: with AI adoption accelerating, governance planning is still well behind.

Visualization of AI governance and browser activity illustrating violations of company AI policies.

48.3% Of U.S. Employees Admit To Breaking Company AI Policies

, , , , , , , , , , , ,

Nearly seven in 10 employees also reported pasting sensitive company data into AI tools, suggesting many organizations face an enforcement problem rather than an awareness problem.

 

Key Takeaways

  • 48.3% of U.S. employees knowingly breach their company’s AI policy.
  • 68.7% said they pasted internal, sensitive, or regulated data into AI tools in the last three months.
  • 70.9% use AI tools daily for work.
  • 42.3% reported using approved and unapproved AI tools side by side during the same browser session.
  • 49.3% said they would use an unapproved AI tool if their preferred tool was not authorized.

Neon Cyber’s report, Quantifying Shadow AI Risk in the Browser, suggests many organizations face a growing AI governance challenge that extends beyond policy development. The survey of 227 U.S. employees who use AI at work found that nearly half knowingly violate their employer’s AI policies despite understanding those rules.

The report argues that the problem is no longer whether organizations have AI policies. Instead, it suggests many organizations struggle to enforce them where employees actually use AI.

Policy is not changing behavior

Among respondents, 63.0% said their organization has a clear AI policy they understand. Of those employees, 48.3% admitted they knowingly breached that policy.

The report also found that 39.6% knowingly used unapproved AI tools, while another 15.0% said they might have. Nearly half (49.3%) said they would simply switch to another unauthorized AI tool if the one they wanted was not approved by their employer.

According to the report, employees often prioritize productivity over internal approval processes, particularly as AI tools become a routine part of daily work.

Sensitive company data continues flowing into AI

The survey found that 68.7% of respondents pasted or uploaded internal, sensitive, or regulated information into AI tools during the last three months, including financial information, customer data, and login credentials.

Shadow AI is becoming routine

More than 70% said they use AI tools daily, while 51.1% reported using them multiple times each day. Nearly two-thirds described AI as essential or very necessary to their jobs.

At the same time, 42.3% said they use both approved enterprise AI tools and unapproved AI services in the same browser session, using the same corporate data. The report argues that this makes it difficult for organizations to detect unauthorized AI activity using traditional security controls.

Browser habits also contributed to the risk. Nearly 70% reported using the same browser profile for both work and personal activities, while 63.0% said they mix personal and work accounts within the same browsing session.

Neon Cyber conducted the survey in May 2026 among 227 self-identified U.S. employees who use AI at work.

Load More