LLMs and Contextual Integrity

I have been thinking a lot about AI and integrity. Part of that is contextual integrity. I recently found two papers on the topic.

CIMemories: A Compositional Benchmark for Contextual Integrity of Persistent Memory in LLMs“:

Abstract: Large Language Models (LLMs) increasingly use persistent memory from past interactions to enhance personalization and task performance. However, this memory introduces critical risks when sensitive information is revealed in inappropriate contexts. We present CIMemories, a benchmark for evaluating whether LLMs appropriately control information flow from memory based on task context. CIMemories uses synthetic user profiles with over 100 attributes per user, paired with diverse task contexts in which each attribute may be essential for some tasks but inappropriate for others. Our evaluation reveals that frontier models exhibit up to 69% attribute-level violations (leaking information inappropriately), with lower violation rates often coming at the cost of task utility. Violations accumulate across both tasks and runs: as usage increases from 1 to 40 tasks, GPT-5’s violations rise from 0.1% to 9.6%, reaching 25.1% when the same prompt is executed 5 times, revealing arbitrary and unstable behavior in which models leak different attributes for identical prompts. Privacy-conscious prompting does not solve this—models overgeneralize, sharing everything or nothing rather than making nuanced, context-dependent decisions. These findings reveal fundamental limitations that require contextually aware reasoning capabilities, not just better prompting or scaling.

Contextual Integrity in LLMs via Reasoning and Reinforcement Learning“:

Abstract: As the era of autonomous agents making decisions on behalf of users unfolds, ensuring contextual integrity (CI)—what is the appropriate information to share while carrying out a certain task—becomes a central question to the field. We posit that CI demands a form of reasoning where the agent needs to reason about the context in which it is operating. To test this, we first prompt LLMs to reason explicitly about CI when deciding what information to disclose. We then extend this approach by developing a reinforcement learning (RL) framework that further instills in models the reasoning necessary to achieve CI. Using a synthetic, automatically created, dataset of only 700 examples but with diverse contexts and information disclosure norms, we show that our method substantially reduces inappropriate information disclosure while maintaining task performance across multiple model sizes and families. Importantly, improvements transfer from this synthetic dataset to established CI benchmarks such as PrivacyLens that has human annotations and evaluates privacy leakage of AI assistants in actions and tool calls.

Posted on August 18, 2026 at 6:40 AM5 Comments

Comments

GregW August 18, 2026 10:51 AM

Regarding “thinking about AI and integrity”, while it depends on what you mean by integrity, in the domain of information processing (not storage itself), the “Integrity” assurances in a deterministic computational system are pretty different from “Integrity” assurances that can be made in a nondeterministic system where many or all steps of data processing occur via nondeterministic LLM agents.

There are ways to request certain models to be deterministic but this issue of nondeterminism gets pretty fundamental with certain types of use cases pretty fast. You start to have to deal with m-of-n voting, distilling and checksumming key outputs, etc to patch over the integrity-of-processing issues (just like with unreliable hardware) or carefully anticipate which steps of a process can or cannot be allowed to be nondeterministic.

ResearcherZero August 18, 2026 11:13 PM

How not to design an algorithm used in an automated assessment tool…

Automated assessment tools do not possess the nuance required for complex assessment of human subjects in real world conditions, living with a complex set of needs in varying situations.

Algorithm was not trialed on real people or developed with input from clinicians.
Errors made by the algorithm compound due to design decisions made during development.

https://www.abc.net.au/news/2026-08-17/inside-the-black-box-aged-care-algorithm-for-support-at-home/107033970

Data on how the algorithm is failing older Australians is not being tracked.
https://www.theguardian.com/australia-news/2026/apr/03/aged-care-funding-assessment-tool-algorithm

Development of the assessment tool failed to adopt good clinical design and error prevention principles during the design process. Principles foundational for sound clinical assessment.

https://doi.org/10.1213/ANE.0000000000003859

lurker August 18, 2026 11:31 PM

@ResearcherZero

“sound clinical assessment” would not have been a parameter for the algorithm. Its purpose would have been to produce a favorable budgetary result. This can be done efficiently by reducing the eligible population.

ResearcherZero August 19, 2026 12:34 AM

@lurker, Bruce

In the original design for The Integrated Assessment Tool (IAT) while it was undergoing testing, it contained an option for human override of assessments. The option is refereed to in the user manual. After the human override option was removed, IAT did not undergo further testing. The removal of this option has led to people who require more care, instead ending up with less care, as a result of errors made by the IAT classification system.

The legality of removing the human override option is questionable and may not hold up under the legislation governing eligibility for Aged Care assistance and support at home.

Given this comes hot on the heels of the Inquiry into Robodebt, the fact that hundreds of requests for review of decisions made by the Integrated Assessment Tool have gone ignored, is deeply troubling. Multiple requests to review the design of the tool also went ignored.

The Robodebt inquiry found that concerns raised during the development of the automated system went ignored by senior public servants and ministers. The Robodebt Royal Commission concluded that the Robodebt scheme was a crude and cruel mechanism, neither fair nor legal.

Two important recommendations made were:

  • more robust accountability measures and regulatory/oversight agencies
  • greater safeguards in relation to privacy, data matching and automation

Another critically important recommendation made by the Royal Commission was:

Restricting the kinds of decisions which can be automated or made without human oversight.

The Australian federal government stated that it would accept all recommendations made “in principle”. Legislation put forward to prevent a repeat of Robodebt is yet to be adopted.

Even worse, the current federal government is poised to revive, what was found to be an illegal automated welfare penalty scheme, without any laws in place to prevent a repeat.

Errors and incorrect assessment of welfare and social services recipients continue to harm vulnerable people, despite government promises that it would implement the recommendations.

https://www.nationaltribune.com.au/advocacy-organisations-oppose-plans-to-revive-harmful-welfare-penalties/

ResearcherZero August 19, 2026 12:39 AM

Liberal senator Paul Scarr looked into the legality of removing human override from IAT.

I did some research and found the guidelines that were referred to in My Aged Care. This is Assessor Portal User Guide 6, and that’s what it says on page 37: it can’t be overridden. It then refers to rule 81-10 of the Aged Care Rules, which make no reference whatsoever to overriding assessments. These Aged Care Rules went before the delegated legislation scrutiny committee, on which I sit. There was no suggestion that there would be these guidelines preventing overruling of assessments.

https://www.openaustralia.org.au/senate/?id=2026-03-11.214.2

Leave a comment

Blog moderation policy

Login

Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via https://michelf.ca/projects/php-markdown/extra/

Sidebar photo of Bruce Schneier by Joe MacInnis.