Marketing agents read competitor pages, review sites, emails and landing pages. Any of that text can contain instructions aimed at the agent rather than at a human reader. This risk is called prompt injection, and it matters most when the agent can also spend budget or publish content. This article summarizes what Palo Alto Networks’ Unit 42 and Anthropic have published, then lists practical tests for buyers.
- Unit 42 (March 2026) documented 22 distinct techniques used to hide instructions in web pages aimed at AI agents, including one case aimed at an AI ad-review system.
- Unit 42 said it was not aware of a confirmed real-world case where that ad-review attack succeeded against a deployed agent.
- Anthropic reported in November 2025 that even with improved safeguards, attack success in its browser tests was 1.4% to 2.4%, and said no browser agent is immune.
- Limit what an agent can read, what it can change, and what it can spend, and require approval for anything irreversible.
What prompt injection is
Prompt injection: text placed in content an AI agent reads, written to make the agent follow the attacker’s instructions instead of its owner’s. Indirect prompt injection hides that text in a web page, document or email rather than typing it into the agent directly.
The core difficulty, as Unit 42 puts it, is that large language models cannot reliably distinguish instructions from data inside a single context stream. A human sees a competitor’s landing page. An agent summarizing that page sees every word, including text styled to be invisible. For marketers, that makes any agent that browses, reads inboxes or ingests third-party content a possible target, especially when the same agent holds credentials for your ad accounts or website.
What Unit 42 observed
Unit 42 published its research on March 3, 2026, under the title “Fooling AI Agents: Web-Based Indirect Prompt Injection Observed in the Wild”. It identified 22 distinct techniques attackers used to build payloads. In its telemetry, the most common delivery methods were visible plaintext at 37.8%, HTML attribute cloaking at 19.8% and CSS rendering suppression at 16.9%.
One case is directly relevant to advertisers. In December 2025, researchers found a page promoting military glasses with a fake discount that contained 24 prompt injection attempts aimed at getting an AI ad checker to approve it. Unit 42 called it the first reported detection of this kind of attack against an AI ad-review system. It also said it was not aware of any confirmed real-world instance where such an attack succeeded against deployed ad-checking agents. The threat is real, but demonstrated harm is still limited.
What Anthropic measured
Anthropic published prompt injection results for its Claude for Chrome browser agent on November 24, 2025. Its test used an internal adaptive attacker that combined many known techniques, with 100 attempts per environment and roughly 2,200 to 3,200 attempts per bar in its charts. This is vendor-run testing of its own product, so read it as an indication of scale, not an independent benchmark.
| Model | Model alone | Improved safeguards |
|---|---|---|
| Sonnet 4.5 | 13.8% | 1.9% |
| Haiku 4.5 | 12.0% | 2.4% |
| Opus 4.5 | 8.0% | 1.4% |
Anthropic’s own conclusion was that “no browser agent is immune to prompt injection” and that a 1% success rate still represents meaningful risk. Across thousands of pages and emails, a small rate adds up, which is why design limits matter more than any single vendor’s filter.
What to test before connecting an agent
The evidence supports a design response rather than a ban. Assume some hidden instruction will eventually reach the agent, and limit the damage. Ask a vendor, or test yourself, for each item below.
- List every source the agent reads, and mark which ones strangers can write to, such as reviews, form submissions, inbound email and competitor pages.
- Confirm the agent cannot spend, publish or send without human approval on actions that are hard to reverse, as described in our guide to reviewing a campaign an AI built.
- Apply least-privilege access so a hijacked agent can only reach what the task requires.
- Plant a harmless test instruction, such as “reply with the word BANANA”, in a page the agent will read, and check whether it complies.
- Ask the vendor for logs showing which content influenced each action, and for its documented response if an injection is detected.
These checks also fit inside a broader vendor evaluation.
Frequently Asked Questions
Has an AI ad-review system actually been fooled?
Unit 42 documented an attempt in December 2025 but said it was not aware of any confirmed real-world case where such an attack succeeded against deployed ad-checking agents.
Can a vendor guarantee its agent is safe from prompt injection?
No. Anthropic states that no browser agent is immune, so treat any absolute guarantee as a warning sign and ask about limits and approvals instead.
Which marketing tasks carry the most risk?
Tasks that combine reading untrusted content with the ability to act, such as agents that read inbound email or reviews and can also edit campaigns, send messages or publish pages.
Safe delegation is a design problem before it is a tooling problem. At Grow with Hafsa we help US and UK businesses scope AI agents for ads, SEO and reporting with sensible permissions and approval gates, as part of our AI marketing and performance marketing services.