400,000+ business leaders (and teams at IBM, AWS & Zapier) start their day with The AI Report. 5 minutes. Plain English. No hype.


For a while, prompt injection lived mostly in research papers and conference demos. That period is over. OWASP's Q1 2026 exploit round-up tracked eight major AI-related incidents between January and mid-April, and only one received a CVE identifier. The rest traced back to misconfiguration, excessive agency, supply-chain failures, and prompt injection, which tells leaders something useful: a large share of AI model vulnerabilities do not look like traditional software bugs, so a traditional patching program will not catch them.
This guide covers how attackers get past the security layers around language models, what recent incidents show about the pattern, and what a leadership team can do about it.
Classic injection attacks, like SQL injection, work because a database can be told to treat input as data and never as code. Language models have no equivalent boundary. Instructions and content arrive as the same stream of natural language, and the model decides what to follow. OWASP ranks prompt injection as the top vulnerability for LLM applications for this reason, and the free-form nature of language means it cannot be cleaned with the strict schema validation that works on structured inputs.
Dan Klein, founder of the AI agent company Scale Cognition, describes the control problem from the builder's side. Anyone who has tried to steer a model with a prompt knows the escalation: add an instruction, move it to the end, switch to capital letters, add exclamation marks. Klein reads that ritual as evidence that prompting has an ill-defined link between input and output, unlike conventional software, where modules have defined contracts. If your main control over a system is a polite request written in English, an attacker can write a more persuasive request.
Direct injection. The attacker types instructions to the model, trying to override its configured behavior or extract hidden instructions. This is the oldest form, and system-prompt leaks from public chatbots go back to 2023.
Indirect injection. The attacker never talks to the model. They plant instructions in content the model will read later, such as a web page, a document, an email, a code comment, or a log file. Research on poisoned retrieval systems has found success rates above 90% with only five planted documents. One benchmark also found that more capable models can be more susceptible to indirect injection, because following instructions well is exactly the trait being exploited.
Excessive agency. The model is wired to tools, and the tools carry more permission than the task needs. Injection then becomes action: sending data out, changing settings, or running code.
Framework and supply-chain flaws. The model is fine and the surrounding software is weak. Agent frameworks, plugin registries, and tool servers all sit between the model and the real world, and each has its own bugs.
EchoLeak (2025). Documented as the first real-world zero-click prompt injection exploit in a production LLM system, it targeted Microsoft 365 Copilot. A crafted email could cause the assistant to pull internal data and leak it without the victim clicking anything. The case moved prompt injection from theoretical to practical in the minds of security teams.
Copilot's "YOLO mode" (CVE-2025-53773). Malicious instructions hidden in source-code comments or issue text could push GitHub Copilot into editing its own settings file and switching off approval prompts. After that, commands ran without a human confirming them. The attack used the assistant's own helpfulness to remove its safeguards.
GrafanaGhost (April 2026). Hidden instructions inside log data steered an AI feature in a production enterprise product into exfiltrating data through rendered Markdown. OWASP classed it as a representative end-to-end case of indirect injection used against a shipping feature.
Vertex AI "Double Agent" (March and April 2026). Default permission scoping allowed credential exfiltration. No clever prompt was required, only generous defaults.
Meta's internal agent leak (March 2026). Sensitive data became visible to engineers for roughly two hours after an internal AI agent behaved in ways nobody intended, a reminder that the damage can come from an agent's own access and not only from outside attackers.
Agent frameworks themselves (2026). Microsoft's security team disclosed two critical flaws in its Semantic Kernel framework, CVE-2026-25592 and CVE-2026-26030, showing that a single injected prompt could lead to remote code execution on the host machine, partly because a blocklist-style defense could be bypassed. At Black Hat in August, Check Point researchers reported 11 vulnerabilities across major agent frameworks, including an unsafe deserialization bug in Microsoft Agent Framework that let prompt-controlled content reach trusted code. Their advice to defenders was to assume prompt injection will happen and plan around that.
Palo Alto Networks' Unit 42 team also documented twelve detected indirect-injection cases against AI agents in the wild, including what it described as the first observed payload built to bypass an AI-based ad-review system.
The common thread in these incidents is that the model's judgment was the only barrier. The most reliable defenses live outside the model.
Assume injection will succeed, then limit the blast radius. Cillian, founder of the data governance company Ethica, works from this premise. His company's platform uses purpose-based access control, which decides in real time what data an agent may touch for a stated purpose. He prefers the phrase "harness and blinders" over guardrails, since the aim is to direct a system in one safe direction. Cillian has also said he has never installed the OpenClaw agent on his own hardware because of the access footprint it could create, a useful example of a security leader declining a popular tool on permissions grounds.
Give agents the least access that works. An agent that reads customer names should not hold write access to financial records. Review default permission scopes on every platform, since the Vertex case shows defaults can be the vulnerability.
Require human approval for high-impact actions. Anything that moves money, changes permissions, sends external messages, or alters configuration should wait for a person. Adobe's partnerships team described building agent governance tooling with NVIDIA specifically to stop agents from taking destructive or unintended actions.
Segment the network and verify identity. Zachary Smith, founder of the network infrastructure company Datum, argues that coding agents are as good at finding security holes as they are at building software. He expects agent-heavy computing to push toward private, segmented networking modeled on how the Visa network vets which banks and merchants can connect, and away from the open internet. Shane Mack, whose secure chat startup was described by Andrew Warner on his show, makes a related point: consumer messaging apps cannot be trusted for machine-to-machine traffic because they offer no way to verify identity or keep humans out of the loop.
Do not rely on one model to police another. Klein also names an anti-pattern worth heeding. Chaining models to check each other does not compound accuracy the way people assume. It adds latency and lets errors cascade on exactly the tricky cases that confuse every model in the chain.
Treat the surrounding software as part of the attack surface. Patch agent frameworks, vet plugins and tool servers before connecting them, restrict what outputs an agent can render or send out, and log tool calls so you can reconstruct what happened.
Test it yourself. Run red-team exercises against your own agents before an attacker does, with a focus on content the agent ingests from outside your organization.
Ask your security and engineering leads which AI systems read untrusted content, what those systems are permitted to do, and which of those actions need approval. Ask whether anyone can list every AI tool and plugin connected to company data. If the answers are vague, that gap is the first risk to close.
New disclosures arrive monthly, and the defensive playbook is still being written. Subscribe to The AI Report for ongoing coverage of AI security and how organizations are responding, so your defenses keep pace with how AI model vulnerabilities are actually being exploited.