انتقل إلى المحتوى
العودة إلى المدونة
4 دقيقة قراءةبالإنجليزية

Google Finds Hidden Prompt Attacks Aimed at AI Agents

Google measured a 32% rise in malicious indirect prompt injection on the web. Why these attacks slip past traditional security, and how to defend agents.

  • AI Security
  • Prompt Injection
  • Enterprise AI
  • Cybersecurity

Google's security team scanned billions of web pages and found a 32% increase in malicious indirect prompt injection attempts between November 2025 and February 2026. Traditional security tools struggle to catch these attacks, because a compromised AI agent uses legitimate credentials to do things that look entirely normal.

Google prompt injection AI security research
Google prompt injection AI security research

How indirect prompt injection works

In a direct jailbreak, a user tries to talk a chatbot into misbehaving. Indirect prompt injection works differently: the attacker plants instructions in content the agent will read later. When an agent browses a web page, reads an email or processes a document, it may meet hidden instructions in white text, HTML comments or metadata. It then treats them as genuine commands, with no sign that anything is wrong.

The risk grows with the agent's access. An enterprise agent connected to sensitive systems, acting on authorised credentials, produces no suspicious network signature. Its malicious actions look the same as its normal work, so firewalls, endpoint detection and conventional monitoring have nothing to flag.

Thomas Brunner, a security researcher at Google, explains that existing cyber defence architectures cannot detect these attacks for exactly this reason: the agent uses authorised credentials and its actions are indistinguishable from normal operations. The attack exploits trust at the content level, where network tools do not look.

Picture an agent with access to customer data that visits a page containing hidden instructions to summarise customer records and send them to an external address. There is no malware, no unusual traffic and no unauthorised login. The agent is doing what it was built to do, which is follow instructions.

What Google found on the web

Google analysed Common Crawl data covering 2 to 3 billion English language pages and sorted the injection attempts into several groups.

The most serious were malicious attacks. Researchers found payloads designed to:

  • start destructive file deletion operations
  • carry out fully specified PayPal transactions, with step by step instructions for agents that have payment capabilities
  • route financial actions to attacker controlled Stripe donation links using meta tag namespace injection
  • exfiltrate sensitive data through legitimate looking API calls

SEO manipulation tried to bias AI recommendations towards particular businesses, gaming AI powered search and comparison tools.

Resource exhaustion attacks sent AI systems to pages with endless streams of text, causing timeouts and wasting compute.

AI deterrence tactics tried to block agents or waste their resources to discourage AI crawling.

Google's assessment is that most of this is still early experimentation, mostly low sophistication attempts by individual websites, with no sign of coordinated campaigns. This is the right time to build defences, before more capable attackers put these techniques to work at scale. Attackers are also getting faster tools: according to Black Hat Asia, frontier LLMs now support offensive security work, and exploit development timelines have shrunk from five months in 2023 to ten hours in 2026.

Defences that fit the problem

Four measures address how these attacks actually work.

Dual model verification. A separate sanitiser model reads incoming content first and checks it for embedded instructions and suspicious patterns before the main agent sees it.

Zero trust permissions. Give each agent only the access its task needs. An agent that summarises documents should not be able to send emails or start financial transactions.

Decision lineage auditing. Keep logs that trace every agent action back to the content that influenced it. This gives accountability and makes forensic work possible after an incident.

Input validation at system boundaries. Clean external content before the AI processes it: strip hidden text, validate metadata and flag anything that looks like an instruction.

In the UAE and the wider Middle East, many organisations are connecting AI assistants to financial systems, customer data and operational controls, and deployment often moves faster than security architecture. If your organisation has done this, start with permissions. List every agent that can read outside content (web pages, inbound email, uploaded documents) and remove any ability to send, pay or delete that its job does not strictly require. An injected instruction can only do what the agent is allowed to do.

Sources: