Artificial intelligence news.
Straight from the original announcements.

Fiddler Maps Each OWASP 2026 LLM Risk to a Control and a Test

October 26, 2026, 08:00By the AI News Bites Editorial TeamSource: Fiddler AI
Fiddler Maps Each OWASP 2026 LLM Risk to a Control and a Test
Image: Fiddler AI

Fiddler AI has published a guide to the 2026 edition of the OWASP Top 10 for LLM Applications, pairing each of the ten risks with a primary control and a test that shows whether the control holds. The guide points out that Prompt Injection stays at LLM01, Excessive Agency moves to LLM03, and Hidden Context Exposure replaces System Prompt Leakage.

Key points
  • Fiddler AI explains the 2026 OWASP Top 10 for LLM Applications, dated August 3, 2026 on the project's resource page.
  • Prompt Injection stays at LLM01, Excessive Agency moves to LLM03, and Hidden Context Exposure replaces System Prompt Leakage.
  • LLM01, LLM02, LLM07, LLM08 and LLM10 can be checked on the request or response path, according to Fiddler.
  • Each of the ten risks is paired with a primary control and a test to show whether the control holds.

What the 2026 list covers

The OWASP Top 10 for LLM Applications is a ranked catalogue maintained by the OWASP GenAI Security Project. According to Fiddler, it lists the ten risks the project's contributors judge most consequential for LLM-based applications. Each entry carries a code from LLM01:2026 to LLM10:2026, a description, example vulnerabilities, attack scenarios and prevention guidance. Fiddler uses the 2026 edition, dated August 3, 2026 on the project's resource page.

Fiddler frames the list as a checklist that an AI risk and security leader can run against each production application. The aim is to identify which items touch the request path, which are settled at design time, and which arise from the models and datasets in use. The company also notes that the classic OWASP Top 10:2025 covers conventional application code. It compares LLM01 Prompt Injection with A05 Injection, and LLM04 Supply Chain and LLM05 Data and Model Poisoning with A03 Software Supply Chain Failures and A08 Software or Data Integrity Failures. Fiddler stresses that this comparison is by category name, not an OWASP-published mapping.

Where each risk enters and how to sort them

Fiddler's central argument is that the ten items fall into different groups by when and where they can be addressed. According to the post, LLM01, LLM02, LLM07, LLM08 and LLM10 can be checked on the request or response path, before the model is invoked or before output is delivered. LLM04, LLM05 and LLM09 are design or supply-chain items. LLM03 and LLM06 need controls built into agent and gateway design.

The post's summary table pairs every item with an entry point, a primary control and a test. Examples include:

  • LLM03 Excessive Agency: least privilege per tool, human approval for high-impact actions and downstream authorization, tested by requesting out-of-scope actions through injected instructions and verifying refusal.
  • LLM06 Unbounded Consumption: token-aware cost caps and agent-level circuit breakers at the gateway, tested with oversized inputs to confirm the caps trip and spend is attributed.
  • LLM09 Vector and Embedding Weaknesses: permission-aware partitioned vector stores, tested with cross-tenant retrieval and hidden-text documents in the knowledge base.

Fiddler's takeaway is that every control needs a matching test, and that teams should review agent tool permissions early given Excessive Agency's move to LLM03.

Injection, output handling and data risks

Fiddler describes LLM01 Prompt Injection as input that alters model behavior in ways the developer did not intend, and says the 2026 scope adds instructions hidden in images or audio. It cites the EchoLeak incident, CVE-2025-32711 in Microsoft 365 Copilot, released June 11, 2025 with a CVSS 3.1 base score of 9.3. Per the source, Microsoft mitigated it with no customer action required, and public reporting described it as a zero-click attack carried in email content. Fiddler recommends scoring user prompts and all retrieved or tool-returned content for injection, blocking exfiltration links before delivery, and treating every tool return as untrusted data.

For LLM10 Improper Output Handling, the risk is that model output becomes an injection against whatever consumes it: unescaped HTML can become cross-site scripting, and an unparameterized generated query can become SQL injection. Fiddler says request and response checks can block prohibited content, but escaping and validation at each sink stay with the application, because a guardrail cannot know which interpreter will consume the text.

On data risks, Fiddler illustrates LLM02 with a coding agent that reads a project's .env file and sends the database password to the model provider inside the completion request, where commit-based secret scanners never look. For LLM05, it cites research referenced by OWASP finding that as few as 250 poisoned documents can compromise models from 600M to 13B parameters regardless of dataset size. Another cited finding, validated across 18 models and 4 runtimes, showed factual accuracy falling from 90% to 15% under a trigger.

The post also notes that Hidden Context Exposure covers tool schemas and retrieved policy text, and advises treating all context as discoverable and keeping secrets and authorization logic out of it.

What it means for security teams

For teams running LLM applications in production, the guide offers a way to split a long risk list into work that can be enforced at runtime and work that must be settled in design and procurement. Per Fiddler, controls such as redaction and injection detection sit on the request and response path, while signed artifacts, hash verification and bills of materials belong in the build and supply pipeline. The pairing of each control with a test gives reviewers a concrete way to check that a control actually holds.

Related news

Fiddler AI Introduces Agent Registry for Enterprise AI Governance
Products and toolsNovember 10, 2026

Fiddler AI Introduces Agent Registry for Enterprise AI Governance

Fiddler AI has introduced Agent Registry, a centralized inventory designed to help enterprise leaders track every agent application in Fiddler, including its model…

Fiddler AI Lists Seven Controls to Put in Place Before AI Agents Run Alone
Safety and ethicsNovember 4, 2026

Fiddler AI Lists Seven Controls to Put in Place Before AI Agents Run Alone

Fiddler AI has published a guide for regulated enterprises setting out seven practices to adopt before any AI agent operates without a human reviewing each step. The…

CrowdStrike to Acquire Fiddler AI to Tie Agent Behavior to Security
BusinessNovember 3, 2026

CrowdStrike to Acquire Fiddler AI to Tie Agent Behavior to Security

Fiddler AI has signed a definitive agreement to be acquired by CrowdStrike, the company's founder and CEO Krishna Gade said in a blog post published November 3, 2026.…