Update LLM01_PromptInjection.md#72
Open
cybershujin wants to merge 1 commit into
Open
Conversation
v2 changes vs. 2026 v1 Description — actor terminology: "operator" → "application developer" for clarity about whose intent is being violated. Anatomy axes — named labels: the three axes are now explicitly named (Delivery surface, propagation behavior, encoding) instead of just described. Axis (a) generalized from "MCP channel" to "tool connection channel" so the taxonomy isn't MCP-specific. Direct / Indirect definitions — small rewordings: "unexpectedly" → "in undesired ways"; "attacker-controlled instructions" → "data which acts as prompt injection" (covers unintentional cases better); "Defenders generally treat" → "Defenders must generally treat." Prevention & Mitigation intro — substantially expanded (the biggest change). v1 had two short paragraphs saying "no fool-proof prevention, go defense-in-depth." v2 replaces this with three paragraphs that: Frame defense as architectural rather than interceptive, citing NIST AI 100-2 E2025, NCSC, and Debenedetti et al. (CaMeL). Connect LLM01 to LLM06 Excessive Agency using the EchoLeak (CVE-2025-32711), Amazon Q, Supabase MCP, and GitHub Copilot (CVE-2025-53773) incidents, and reference Simon Willison's "lethal trifecta" as a pre-deployment check. Distinguish controls that reduce injection success from controls that bound blast radius, and flag capability-budgeting (Controls 4, 11) as load-bearing for agentic deployments. Control 4 — procurement framing added: new paragraph treating the deterministic policy engine as a procurement/design-review baseline expectation, with citations to NIST AI 100-2 E2025 and the CISA + Five Eyes OT joint guidance (Dec 2025). Known-limits and citations updated accordingly to flag that contractual requirements without verification degrade to the same convenience-permission failure mode. No changes to: Common Examples of Vulnerability (1–8), Controls 1–3 and 5–11 body text, Example Attack Scenarios (1–9), Reference Links, or Related Frameworks list. Signed-off-by: Rachel James <162050062+cybershujin@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
v2 changes vs. 2026 v1
Description — actor terminology: "operator" → "application developer" for clarity about whose intent is being violated.
Anatomy axes — named labels: the three axes are now explicitly named (Delivery surface, propagation behavior, encoding) instead of just described. Axis (a) generalized from "MCP channel" to "tool connection channel" so the taxonomy isn't MCP-specific.
Direct / Indirect definitions — small rewordings: "unexpectedly" → "in undesired ways"; "attacker-controlled instructions" → "data which acts as prompt injection" (covers unintentional cases better); "Defenders generally treat" → "Defenders must generally treat."
Prevention & Mitigation intro — substantially expanded (the biggest change). v1 had two short paragraphs saying "no fool-proof prevention, go defense-in-depth." v2 replaces this with three paragraphs that:
Frame defense as architectural rather than interceptive, citing NIST AI 100-2 E2025, NCSC, and Debenedetti et al. (CaMeL). Connect LLM01 to LLM06 Excessive Agency using the EchoLeak (CVE-2025-32711), Amazon Q, Supabase MCP, and GitHub Copilot (CVE-2025-53773) incidents, and reference Simon Willison's "lethal trifecta" as a pre-deployment check.
Distinguish controls that reduce injection success from controls that bound blast radius, and flag capability-budgeting (Controls 4, 11) as load-bearing for agentic deployments.
Control 4 — procurement framing added: new paragraph treating the deterministic policy engine as a procurement/design-review baseline expectation, with citations to NIST AI 100-2 E2025 and the CISA + Five Eyes OT joint guidance (Dec 2025). Known-limits and citations updated accordingly to flag that contractual requirements without verification degrade to the same convenience-permission failure mode.
No changes to: Common Examples of Vulnerability (1–8), Controls 1–3 and 5–11 body text, Example Attack Scenarios (1–9), Reference Links, or Related Frameworks list.