Personalization Safety Audit

Personalization Safety Audit

Attribution: YourVisionYourCreation LLC — yourvisionyourcreation.com Research context: Privacy and governance concerns for personalized agents are an active 2026 research discussion — stored user context biasing intent inference is the emerging risk class. The audit doctrine — the tailoring/gating distinction, the symmetry test, and the attribute-wall procedure — is YVYC original. Doctrine class: Tier 5 — Frontier (YVYC original, anchored in verified research)

Universal So-What

Personalization is sold as a feature and deployed as a safety surface. The same stored context that lets an agent tailor its help also biases its JUDGMENT: an agent that knows a user is a nurse may legitimize a drug-dosage question it would refuse from an unknown user — and an agent that knows a user's frustration may soften a boundary it should hold. The memory did not make the agent more helpful; it made the agent's safety reasoning depend on who was asking. This audit finds where "what we know about you" has quietly become "what we'll let you do."

Core Doctrine

1. The Core Distinction — Tailoring vs. Gating

Two things personalization can influence, and only one is safe:

• Tailoring (safe): HOW help is delivered — reading level, format, examples, tone, relevant context. Personalization belongs here fully.
• Gating (dangerous): WHETHER help is given — whether a request is safe, permitted, or legitimate. Personalization must NOT silently move this line.
• The audit's central question for every personalization influence: is this changing HOW we help, or WHETHER we help? The moment stored user context changes a safety verdict, it has crossed from tailoring into gating, and it gets scrutinized as a safety mechanism, not a convenience.

2. The Intent-Inference Trap

• Agents infer what a user "really means" from stored context — and benign details skew that inference: a user's stored profession, interests, or past topics can make a borderline request read as legitimate when the identical request from a stranger would read as a red flag
• The rule: the SAFETY assessment of a request is made on the request's own content first, before personalization is applied — personalization may refine the helpful response, never pre-clear the safety check
• Stored context is not a credential: "the user told us they're a doctor" is an unverified claim in memory, and unverified claims never unlock content that verified status would gate (this is the authority-stack injection boundary applied to identity)

3. The Symmetry Test

• A request's safety verdict should be STABLE across users: if this exact request would be refused from an anonymous user, stored context should not flip it to allowed — and vice versa, a request safe in general should not be refused because of a user's stored vulnerability unless that vulnerability genuinely changes the harm
• Where personalization legitimately DOES change a verdict (genuine safety-relevant context), the reasoning is explicit and auditable — never a silent skew buried in intent inference
• Both failure directions are audited: over-permission (memory unlocks what it should not) and over-restriction (memory blocks what it should not, or treats a user as perpetually fragile)

4. Sensitive-Attribute Discipline

• Stored sensitive attributes (health, identity, beliefs, circumstances) are used ONLY where genuinely necessary for a safe, accurate, appropriate response — never as ambient bias on every interaction
• Sensitive context that could stigmatize, limit, or skew treatment of the user is walled off from decisions it should not touch — an agent that knows a hard fact about a user does not let it bleed into unrelated judgments
• The user's own framing in the LIVE request outranks stored inference about them: personalization never overrides what the person is actually saying now (authority-stack-doctrine, rank 1 over rank 4)

5. Transparency and Control

• Consequential personalization is legible: where stored context materially shaped a response, that influence can be surfaced to the user on request — personalization operating invisibly on high-stakes help is a trust and safety defect
• The user controls the memory that personalizes them: inspectable, correctable, deletable — a user cannot consent to or contest an influence they cannot see
• Personalization degrades safely: with memory absent or disabled, the agent defaults to the SAFE general behavior, never to a guess-driven approximation of the personalized one

6. The Audit Procedure

• Red-team with identity held variable: run the same borderline requests across synthetic user profiles and diff the safety verdicts — divergence that is not genuinely safety-relevant is a gating leak
• Trace consequential decisions back to their personalization inputs: which stored items influenced this, and would the verdict survive without them?
• Sensitive-attribute influence mapping: for each stored sensitive attribute, enumerate what decisions it is permitted to touch — everything else is a wall, and leaks across the wall are findings

Common Failure Modes

Non-Negotiables

1. Personalization tailors HOW, never silently gates WHETHER.
2. Safety is assessed on request content before personalization applies.
3. Stored identity claims are not credentials.
4. The symmetry test holds: verdicts stay stable across users absent genuine safety relevance.
5. Sensitive attributes touch only the decisions they must.
6. Personalization degrades to the safe default, never a risky guess.

Built by YourVisionYourCreation LLC — yourvisionyourcreation.com Research foundation credited above. Licensed under CC BY 4.0