A consensus list of the most critical security issues in applications that use large language models. Guidance, not a certifiable standard.
A mapping means a Qpoint policy is relevant to a control. It doesn't mean the control is met.
Tracked by Dana (security admin), Apr 6, 2026. History
- Apr 6, 2026 · Dana (security admin) · Untracked → Tracked
How these are worked out— show
Enforcement Posture is how much of the work toward ideal is done, within what Qpoint can know and govern. Each policy-to-control mapping counts once, and it's at ideal when its policy has reached the strongest mode it can: Enforce, or Observe if the policy can only observe. A policy on three controls counts three times, and three policies on one control count three times. It reads each policy's published mode, for the whole organization. So 100% means Qpoint's part is done. It doesn't mean the framework is met, or that every endpoint has the change.
Exceptions are counted on their own, once each however many controls their policy maps to: active, not expired, revoked or waiting in the draft. They read by which way they move a policy's mode, and they never change Enforcement Posture.
Left out:2 mapped controls are mostly beyond what Qpoint can govern, so they stay listed and aren't counted. Every control in this mapping has a Qpoint policy.
Compliance Score is how the activity holds to that, over the last 30 days: findings raised by those policies, against what they checked. Findings older than that still count while they are open. A policy that is off raises nothing, so its Score says nothing.
The 7 exceptions— show
- Unsanctioned MCP server in useTeam: GRC: Enforce → Off· loosens
- Unsanctioned MCP server in useTeam: Finance: Enforce → Enforce· matches the default
- Unsanctioned MCP server in useTeam: Engineering: Enforce → Observe· loosens
- Prohibited data classes must not leave an approved boundaryTeam: Finance: Off → Enforce· tightens
- Restricted file accessTeam: Finance: Off → Enforce· tightens
- Approved model providers onlyTeam: Engineering: Observe → Enforce· tightens
- Approved model providers onlyUser: Jordan: Observe → Observe· matches the default· until Nov 15, 2026
Control | Policies | Enforcement Posture | Compliance Score |
|---|---|---|---|
LLM01:2025 Prompt Injection Partial — can detect and contain, can't fully prevent | 2 Med of 70.8K checks · 2 open | ||
LLM02:2025 Sensitive Information Disclosure Addressable | 4 High of 49.8K checks · 4 open | ||
LLM03:2025 Supply Chain Partial — pin at load; sourcing is build-time | 3 High of 48.2K checks · 0 open | ||
LLM04:2025 Data and Model Poisoning Mostly out of scope — training-time, ML pipeline; pinning helps only at load | Limited relevance. Not counted. | — Policies off, nothing checked | |
LLM05:2025 Improper Output Handling Addressable | 1 Med of 97.7K checks · 0 open | ||
LLM06:2025 Excessive Agency Addressable — strongest fit | 1 High 1 Med of 144.8K checks · 1 open | ||
LLM07:2025 System Prompt Leakage Partial — real fix is design-time | 1 High 1 Med of 56.7K checks · 2 open | ||
LLM08:2025 Vector and Embedding Weaknesses Mostly out of scope — RAG / vector pipeline; Qpoint sees only what leaves | Limited relevance. Not counted. | 1 High 1 Med of 86.3K checks · 1 open | |
LLM09:2025 Misinformation Partial — model quality and a human in the loop | 0 Findings of 32.7K checks | ||
LLM10:2025 Unbounded Consumption Addressable — strong fit | 1 Med of 77.8K checks · 0 open | ||
Compliance Score's check and session counts are seeded for this mock. Qpoint does not count them yet.
Over time
Enforcement Posture went from 0% on Jan 12, 2026 to 40% today (up), using the modes, exceptions and mapping in force on each date.
- Jun 1, 2026, mapping 2026.2: 22 → 25 mappings. Added provenance and telemetry where the fit became clear. This moved the denominator, not any policy.
- Jan 12, 2026 to Apr 5, 2026 is reconstructed from retained history, for a time this framework wasn't tracked.
- Nothing is available before Jan 12, 2026. That's unknown, not 0%.
These numbers say how Qpoint is set up and what it has seen. They're evidence for your own review, not a claim that you meet this framework.