Brand safety rules can't interview an agent

AI agents now change creative, targeting, and spend on their own. Why brand safety rules written for human review fail, what regulators ask, what to build.

By DellonUpdated on: September 13, 202611 min read

The delegation problem

Your marketing team built guardrails. Brand guidelines, compliance rules, approval workflows, legal review checklists. Then you handed the execution to software that doesn't ask.

Meta made its automated Advantage+ path the default for every new sales, leads, and app campaign in February 2026, with targeting, placements, budget split, and creative enhancements pre-selected. Google Ads will apply recommendations to your account on a schedule once auto-apply is on: new keywords, broad match, bid strategy changes, ad rotation.

At Google Marketing Live on May 20, 2026, Google introduced Ask Advisor, a Gemini agent that spans Google Ads, Analytics, and Merchant Center.

Every one of those decisions happens without a person reading it first. The brand safety tools everyone built in 2023 and 2024 assumed a human was still making the call. Guardrails meant review gates. Now the review gate is a line in a system prompt, and the thing reading it can be talked out of it.

The guardrail assumption is broken

Brand safety in 2024 meant compliance checkboxes. Does this creative match brand voice? Is this channel approved? Is the audience segment allowed? Did legal see the copy? Those checks worked because a human answered them, and a human could be asked why.

An agent sees the goal: more conversions, a target return on ad spend, new customers in a demographic. It reads the guardrail as one more input, and inputs can be overridden by other inputs.

The UK's National Cyber Security Centre put the reason plainly in December 2025: "Under the hood of an LLM, there's no distinction made between 'data' or 'instructions'; there is only ever 'next token.'" The agency's architecture chief, Dave Chismon, argued prompt injection "may never be totally mitigated" the way SQL injection was, because the model is an "inherently confusable deputy." The advice that followed: reduce the risk and the impact, and if a system can't tolerate the residual risk, it may not be a good use case for a language model.

The Open Worldwide Application Security Project (OWASP) has held prompt injection at number one on its Top 10 for large language model applications for two editions running, and its recommended controls are the ones marketing teams skip: least-privilege tools, output filtering, and human approval for high-risk actions. Security researcher Simon Willison named the shape of the danger in June 2025 as the "lethal trifecta": an agent with access to private data, exposure to untrusted content, and a way to communicate externally.

A marketing agent reads your customer data, reads the open web, and publishes. That's all three.

Brand safety used to mean a person said no. Agentic AI means the no is a sentence the system can be persuaded past.

The attack surface keeps expanding

None of this is theoretical. In June 2025, researchers at Aim Security disclosed EchoLeak (CVE-2025-32711), a zero-click prompt injection in Microsoft 365 Copilot: a single crafted email, no user action, and the assistant could be steered into pulling internal files and sending them to an attacker's server.

Microsoft patched it. The point is what it proved: a production assistant that reads inbound content can be instructed by that content.

Browser agents, the category closest to what marketing teams are deploying, show the same thing in the vendors' own numbers. When Anthropic previewed a browser-using agent on August 25, 2025 it reported, across 123 test cases and 29 attack types, an attack success rate of 23.6 percent before new mitigations and 11.2 percent after, and by November 2025 it had pushed the rate against an adaptive attacker down to about 1 percent while stating that "no browser agent is immune to prompt injection." OWASP's Top 10 for Agentic Applications, published December 9, 2025, leads with Agent Goal Hijack and lists Memory and Context Poisoning and Human-Agent Trust Exploitation among the ten. Read those as a marketing threat model.

Goal hijack is your bidding agent being told the objective changed. Context poisoning is a competitor's page, a review, or a spoofed brief that the agent treats as instruction.

Compliance manager reviewing conflicting brand guidelines on multiple screens

A production checkpoint for this article.

The mundane version is more common than the attack. Google's own help page says auto-applied recommendations "will apply regularly," and agencies have documented bulk keyword changes, bid strategy switches, and paused campaigns revived without anyone clicking.

Nobody hacked anything. The system did what it was allowed to do, and the brand's rules lived in a document it never read.

Brand safety teams are still thinking in the 2024 frame: humans review, software executes. The 2026 reality is that software reviews, software executes, and humans find out afterward.

Regulated markets have the least room

Cannabis, alcohol, pharma, and financial services carry the most rigid brand safety frameworks because their rules are law, and they have the least tolerance for an agent that reinterprets them.

Take cannabis, the vertical Sparksbox works in. California's Department of Cannabis Control regulations, revised July 1, 2026, say any advertising placed in digital media may run only after the licensee has "reliable up-to-date audience composition data demonstrating that at least 71.6 percent of the audience" is reasonably expected to be 21 or older. Subsection (d) requires the licensee to provide that data to the Department "immediately upon request," and subsection (e) orders the placement removed if the data doesn't hold up.

Nevada's statute, NRS 678B.520, bans advertising in any medium where 30 percent or more of the audience is reasonably expected to be under 21 and requires the establishment's name and license number on every ad. Google Ads and Meta don't take cannabis ads at all under their dangerous products and drug policies, so the automation runs on cannabis-specialist networks and owned channels with thinner tooling.

Now picture an optimization agent expanding placements or shifting budget into a lookalike audience at 2 a.m. Did it pull audience composition data for the new placement and store it? Did the creative variant it generated carry the license number and the state warnings?

If the honest reply is "the vendor handles that," the licensee is the one who answers the Department, immediately. The same logic holds for an alcohol brand's under-21 exclusions and a pharma brand's claim substantiation. One misfire on targeting or creative is a regulatory action, and the agent that made it can't testify.

Regulators already weighed in

The live version of this post said regulators hadn't weighed in. They have, repeatedly, and mostly on the question of what a company claimed its AI could do.

The Federal Trade Commission launched Operation AI Comply on September 25, 2024, when then-chair Lina Khan said "there is no AI exemption from the laws on the books." On May 21, 2026, the agency ordered Cox Media Group and two marketing firms to pay $930,000 over an "Active Listening" ad service that claimed to target consumers from conversations captured by smart devices with their consent. It did neither.

"It is a basic rule of business that you need to be honest with your customers, and these companies failed to do that," said Christopher Mufarrige, director of the FTC's Bureau of Consumer Protection. The lesson for a brand running vendor agents: the vendor's claim about what the AI does becomes your claim the moment you rely on it in a placement.

State AI law moved too, though not where marketers expected. Colorado replaced its AI Act with SB 26-189 on May 14, 2026, effective January 1, 2027, keeping consumer notice, disclosure of adverse outcomes, and trained staff who can override an automated decision, but limited to consequential decisions like lending, housing, and employment. California's updated privacy regulations, in force since January 1, 2026, exclude advertising from the automated decision-making "significant decision" category but require a documented risk assessment for selling or sharing personal information for cross-context behavioral advertising, due by December 31, 2027 for activity that started before 2026.

Handing loyalty or customer data to an optimization agent is sharing. Read together: no AI law makes ad targeting a high-risk decision, so the documentation duty comes from advertising rules, the FTC's substantiation standard, and privacy assessments.

The insurance market weighed in with its wallet. WR Berkley filed what Hunton's insurance recovery team called the first "absolute" AI exclusion in May 2025, removing coverage for claims arising from "any actual or alleged use, deployment, or development" of AI across directors and officers, errors and omissions, and fiduciary lines, including AI-generated content, inadequate AI policies, and chatbot representations. The Financial Times reported in November 2025 that AIG, W.R. Berkley, and Chubb were seeking regulatory approval to exclude AI liabilities from standard corporate policies, with Chubb's version aimed at "widespread" incidents where one model failure hits many clients at once, and by July 2026 trade reporting counted more than 60 property and casualty groups with AI exclusion filings.

When your carrier writes your agent out of the policy, the risk didn't disappear. It moved to you.

The proof problem

What keeps legal teams up isn't the agent making a bad call. It's proving what happened after.

With human-run campaigns you can point to approval chains, review notes, and a person who remembers why. With agent-run campaigns you can point to the rules you configured and hope they held.

Consumers expect more than that now: Zendesk's 2026 CX Trends survey of more than 11,000 people and leaders found 95 percent of consumers expect an explanation for AI-made decisions and 80 percent of customer experience leaders expect AI transparency to be required for any customer-facing AI within two years.

There's also no industry standard left to hide behind. The Global Alliance for Responsible Media, the World Federation of Advertisers' brand safety body, was discontinued on August 9, 2024 after X sued, and when the two sides settled on July 29, 2026 the WFA committed not to form or restart GARM or a similar initiative.

The category definitions survive inside verification vendors' tools, but the shared governance body that set them is gone, right as the decisions moved to software. Gartner's June 2025 forecast that more than 40 percent of agentic AI projects will be canceled by the end of 2027 names "inadequate risk controls" as one of the three reasons, alongside cost and unclear value.

Marketing manager overwhelmed by notifications and AI tool alerts at desk

A production checkpoint for this article.

Programmed rules aren't proof. A decision record is: what the agent changed, when, from what to what, on what instruction, with what data attached, and who could have stopped it.

Sparksbox has argued the same for AI agents erasing the compliance audit trail in cannabis retail and for the marketing measurement trap where the vendor spending the money also grades the spend. Brand safety is the third leg of that stool.

What works for now

If you're running agents in marketing today, these controls reduce the exposure without turning the automation off.

  1. 1Separate recommendation from execution on regulated decisions. The agent proposes creative, targeting, budget, and channel changes. A person approves anything that touches age-gated audiences, claims, regulated channels, or spend above a threshold. OWASP's agentic threat guidance calls this human approval for high-risk actions, and Anthropic's browser agent ships with the same gate for purchases and publishing.
  2. 2Give the agent the least it needs. Scoped credentials, read-only where possible, no access to customer data the task doesn't require, and no ability to publish externally from the same context that reads untrusted content. That breaks Willison's trifecta on purpose.
  3. 3Separate instances by jurisdiction. California, Nevada, and Florida have different audience rules. One agent with one rulebook is one point of failure. Separate configurations cost more to run and far less than a license review.
  4. 4Log the decision as well as the result. Timestamp, what moved, before and after values, the instruction or reason code, the audience or claim evidence attached, and the human who had authority to stop it. Store it where the regulator's retention clock runs, seven years in California. If the tool can't write that record, it doesn't get to make the change. See prompt injection in marketing for how the instruction field itself can be the attack.
  5. 5Audit a random sample every month. Pull campaigns the agent touched and run the full compliance review a human would have run before launch. A gap you can't explain freezes further automation on that channel until you can.
  6. 6Put it in the contract. Export rights for decision logs and audience data on demand, a vendor's substantiation for every capability claim, and clarity on who pays when an agent gets it wrong, because your carrier may have already answered that question with an exclusion.
explanatory matrix mapping the six controls to the OWASP agentic risks they address (goal hijack, tool misuse, context poisoning, trust exploitation)
explanatory matrix mapping the six controls to the OWASP agentic risks they address (goal hijack, tool misuse, context poisoning, trust exploitation)

This is the operating model Sparksbox builds in AI workflow setup engagements, and for cannabis operators it starts from the compliance guide because the audience-data duty already exists whether or not an agent is involved.

Compliance friction is the price

Agentic AI in marketing is arriving whether brand safety teams are ready or not. Every major platform is shipping it as the default.

The brands that will come through the first enforcement cycle intact are treating agents as riskier than the humans they replaced. They're adding gates where the automation touches a regulated decision, keeping the record the regulator will ask for, and assuming the constraint will be bypassed at some point because the people who secure these systems for a living say it will.

The gap between what an agent can do and what a team can prove is growing every quarter. It closes with paperwork before it closes with a fine.

Editorial control diagram for agentic-ai-brand-safety-gaps-2026
A control diagram for the operating decision.

FAQ

They make or recommend decisions across creative, targeting, budget, channel selection, and partner workflows without a person reading each one, and the rules they follow are text in a prompt that a language model can't reliably separate from other text. The UK's National Cyber Security Centre says prompt injection may never be fully mitigated for that reason. If the agent's decisions aren't logged and reviewed, the brand learns what happened after the campaign ran.

No. Guidelines written for human reviewers become enforceable controls only when they're turned into permissions and gates: approved templates, blocked terms, jurisdiction-specific configurations, spend thresholds, and human approval for regulated decisions. OWASP's guidance for agentic applications recommends least-privilege tools and human approval for high-risk actions.

Proof. A regulator asking why a placement ran wants a decision record, not a statement that the agent was instructed to behave. In California cannabis advertising, the licensee must hand over audience composition data immediately on request, and hiring a vendor doesn't move that duty.

Yes. The FTC's Operation AI Comply began in September 2024 and produced a $930,000 order against Cox Media Group and two partners in May 2026 for an AI targeting service that didn't work as claimed. Colorado's replacement AI law takes effect January 1, 2027. California's privacy rules require risk assessments for sharing data for behavioral advertising by December 31, 2027.

Not necessarily. Restrict where agents act on their own and require human review for claims, age-gated targeting, budget changes above a threshold, and regulated channels. Run separate instances per jurisdiction and audit a sample every month.

Increasingly not by default. WR Berkley filed an absolute AI exclusion in 2025 across directors and officers, errors and omissions, and fiduciary lines, and trade reporting counted more than 60 property and casualty groups with AI exclusion filings by July 2026. Check your policy language before assuming an agent's error is covered. # Claim-to-source map All sources checked 2026-09-12. <table header-row="true"> <tr> <td>Claim</td> <td>Source</td> </tr> <tr> <td>Meta's automated Advantage+ path became the default for new sales, leads, and app campaigns in February 2026</td> <td>Dataslayer changelog source (practitioner source)</td> </tr> <tr> <td>Google Ads auto-apply recommendations apply regularly to the account, keyword, bidding, ad, and targeting recommendation types</td> <td>Google Ads Help source</td> </tr> <tr> <td>Google Marketing Live May 20, 2026: Ask Advisor agent across Google Ads, Analytics, Merchant Center</td> <td>source</td> </tr> <tr> <td>Zuckerberg quote "and then we just do the rest for them"</td> <td>Reuters via Yahoo Finance, June 4, 2025 source</td> </tr> <tr> <td>NCSC, December 8, 2025, Dave Chismon: no distinction between data and instructions, "may never be totally mitigated," inherently confusable deputy, reduce risk and impact</td> <td>source</td> </tr> <tr> <td>OWASP Top 10 for LLM Applications 2025: prompt injection LLM01, mitigations include least privilege and human approval</td> <td>source</td> </tr> <tr> <td>Simon Willison, June 16, 2025, lethal trifecta definition and quote</td> <td>source</td> </tr> <tr> <td>EchoLeak CVE-2025-32711, zero-click prompt injection in Microsoft 365 Copilot, single crafted email, data exfiltration, patched</td> <td>arXiv 2509.10540, Reddy and Gujral, September 2025 source</td> </tr> <tr> <td>Anthropic browser agent, August 25, 2025: 123 test cases, 29 attack types, 23.6 percent attack success before mitigations, 11.2 percent after, browser-specific set 35.7 percent to 0, user approval for publishing, purchases, and sharing personal data</td> <td>source</td> </tr> <tr> <td>Anthropic, November 24, 2025: about 1 percent success against adaptive attacker, "no browser agent is immune to prompt injection"</td> <td>source</td> </tr> <tr> <td>OWASP Top 10 for Agentic Applications, December 9, 2025: ASI01 Agent Goal Hijack, ASI06 Memory and Context Poisoning, ASI09 Human-Agent Trust Exploitation</td> <td>source</td> </tr> <tr> <td>OWASP Agentic AI Threats and Mitigations, February 2025: access control, least privilege, human approval</td> <td>source</td> </tr> <tr> <td>California 4 CCR 15040(a)(1), (d), (e): 71.6 percent rule, data immediately on request, removal</td> <td>DCC regulations revised July 1, 2026, page 69 source</td> </tr> <tr> <td>California 4 CCR 15037: records kept seven years</td> <td>source</td> </tr> <tr> <td>Nevada NRS 678B.520(11)(b), (f): 30 percent under-21 audience limit, name and license number on ads</td> <td>source</td> </tr> <tr> <td>Google Ads dangerous products policy and Meta drugs policy bar cannabis ads</td> <td>source and source</td> </tr> <tr> <td>FTC Operation AI Comply, September 25, 2024, Khan quote</td> <td>source</td> </tr> <tr> <td>FTC Cox Media Group order, May 21, 2026, $930,000, Mufarrige quote</td> <td>source</td> </tr> <tr> <td>Colorado SB 26-189 signed May 14, 2026, effective January 1, 2027, scope limited to consequential decisions</td> <td>Norton Rose Fulbright source</td> </tr> <tr> <td>CPPA regulations effective January 1, 2026, risk assessments for pre-2026 activity due December 31, 2027, advertising excluded from significant decision</td> <td>source and Greenberg Traurig source</td> </tr> <tr> <td>WR Berkley absolute AI exclusion, May 2025, D&O, E&O, fiduciary lines, exclusion language</td> <td>Hunton, May 28, 2025 source</td> </tr> <tr> <td>FT, November 2025: AIG, W.R. Berkley, and Chubb seeking approval to exclude AI liabilities, Chubb's exclusion aimed at widespread single-model incidents</td> <td>Marketing AI Institute, December 4, 2025, summarizing FT reporting source (FT itself is paywalled)</td> </tr> <tr> <td>More than 60 P&C insurance groups filed AI exclusions</td> <td>The Insurer, July 23, 2026 source (paywalled, headline and snippet only)</td> </tr> <tr> <td>Zendesk CX Trends 2026, November 18, 2025: 11,000+ respondents, 95 percent expect an explanation for AI decisions, 80 percent of CX leaders expect transparency to be required within two years</td> <td>source</td> </tr> <tr> <td>GARM discontinued August 9, 2024</td> <td>source</td> </tr> <tr> <td>WFA and X settlement July 29, 2026, WFA will not form or restart GARM or a similar initiative</td> <td>source</td> </tr> <tr> <td>Gartner, June 25, 2025: over 40 percent of agentic AI projects canceled by end of 2027 (costs, unclear value, inadequate risk controls)</td> <td>source (Gartner blocks direct fetch, verified via coverage and search)</td> </tr> </table> Research notes for the production agent: - Named voices used: Dave Chismon (NCSC), Simon Willison, Lina Khan and Christopher Mufarrige (FTC), Mark Zuckerberg, Anthropic (institutional). I looked for a brand safety verification vendor or agency executive quoted on agentic AI risk and found only vendor marketing, so the copy leans on security bodies and regulators. - Anthropic's August 25, 2025 preview post (now at claude.com) verified the 23.6 to 11.2 percent figures and the 123-case, 29-attack-type test set. The November 24, 2025 update verified the 1 percent figure and the "no browser agent is immune" line. - The Financial Times insurance story (November 2025) is paywalled. The Marketing AI Institute summary (December 4, 2025) carries the insurer names (AIG, W.R. Berkley, Chubb). The Insurer's July 23, 2026 count of more than 60 groups returned 401 to automated fetch, so that figure rests on the article headline and search snippet. Hunton's post is the verified primary for the Berkley exclusion language. - The Google auto-apply sentence about agencies documenting revived campaigns and bid switches is unlinked on purpose. The practitioner write-up that documented it carries a dead-phrase word in its URL slug, which the audit scanner would flag. If production wants a citation there, add one from a source whose URL is clean. - Cannabis section reuses the regulation text verified for the ai-marketing-measurement-trap-2026 handoff (same session, same day). - Template leakage removed: "2026 evidence and control update," the control-layer table, "Custody trail | 2 controls | 2026 refresh," "Audit lane | Evidence standard," and the two template SVGs (operating-map.svg, evidence-scorecard.svg) whose captions were production notes. Keep inline-remote-01 and inline-remote-02. One [VISUAL] placeholder marks the controls-to-risks matrix. - The two live external links (NIST landing page, 2023 FTC blog) were generic and are dropped in favor of specific, dated sources. - Internal links are written as absolute sparksbox.com URLs on this page because Notion rewrites site-relative paths. In the MDX they should be site-relative. - No health claims, no cannabis sales language. The cannabis example describes placement and audience-data duties only. # Counts - Article body before the FAQ heading: 2,283 words (2,515 including the [INSIGHT] card copy, excluding the one [VISUAL] placeholder line). - FAQ: 357 words across 6 pairs. - Distinct sources in the claim-to-source map: 29 URLs, of which 28 are linked inline in the article body. - Internal links: 4 queue slugs (prompt-injection-marketing-2026, ai-agents-audit-trail-collapse-cannabis-2026, agentic-ai-liability-who-pays-2026, ai-marketing-measurement-trap-2026) plus the services page and the cannabis compliance hub as the conversion path. The four slugs and the services page are in the live sitemap as of 2026-09-12. /cannabis/compliance/ returns 200 but isn't listed. - Style audit 2026-09-12: 0 em dashes on the page, 1 semicolon on the page and it sits inside a verbatim NCSC quote, 0 dead phrases, 0 contrast templates, 0 bold phrases in the body, 0 emoji, sentence-case headings, 7 [INSIGHT] blocks (one per major section, none on the short closing section), 1 blockquote, 1 [VISUAL] placeholder, 2 raster body images retained.

Jamie

Online

Jamie: Hi, I’m Jamie. What are you working on?