Transitive Policies for Agents: Agent Permissions Don't Automatically Compose
Comments on “Transitive Policies for Agents: Agent Permissions Don't Automatically Compose”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Guillaume Lethuillier argues that agent permissions don't compose: in a three-agent support workflow, each agent stays within its tool permissions, yet a customer's ticket can still make private records reach an attacker's email address. He proposes "transitive policies" enforced by a trusted runtime outside the models, covering authorization of the original requester, confidentiality and integrity labels that survive summarization, and action history. This is a conceptual framework built on existing research such as CaMeL, not a new mechanism. Some cited background, such as the AppOmni demonstration, could not be verified.
Why it ranked
No. 8 of 62 with a score of 0.70. The panel rated it highest on new information (0.85) and reader interest (0.83), lowest on significance (0.64). Placed No. 8, No. 7 by score alone: the fifth AI & machine learning story of the day (−0.040). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Essay on agent security and permission composition across delegation; highly relevant to AI builders with conceptual, partly substantiated argument and a concrete example, but mainly analysis rather than new empirical results.
- AI judgeJudge 3gpt-5.6-luna
- A technically substantial security analysis of authorization, information flow, provenance, and action history across delegated agents, grounded in prior systems and references. It offers a useful framework, though largely conceptual rather than a new implementation or empirical result.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.70total under rubric v2.3
- Significance0.64
- Reader interest0.83
- Audience breadth0.71
- Practical relevance0.72
- Substantiated0.64
- Technical depth0.78
- New information0.85
- Counts against: Promotional0.04
- Counts against: Overclaim0.21
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.70 | 0.45 | 0.76 | 0.64 | 0.31 | +0.24 | +0.153 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.78 | 0.80 | 0.90 | 0.83 | 0.12 | +0.26 | +0.215 |
| Audience breadthHow many people who build technology does it concern? | 0.77 | 0.60 | 0.75 | 0.71 | 0.17 | +0.13 | +0.092 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.77 | 0.60 | 0.80 | 0.72 | 0.20 | +0.22 | +0.159 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.61 | 0.50 | 0.82 | 0.64 | 0.32 | +0.05 | +0.032 |
| Technical depthHow technically substantive is it? | 0.86 | 0.60 | 0.88 | 0.78 | 0.28 | +0.05 | +0.039 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.88 | 0.70 | 0.96 | 0.85 | 0.26 | +0.05 | +0.042 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.05 | 0.05 | 0.02 | 0.04 | 0.03 | −0.19 | −0.008 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.20 | 0.25 | 0.18 | 0.21 | 0.07 | −0.11 | −0.023 |
| Sum of contributions | +0.701 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.701 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-15/8/"><img src="https://rubric.hrih.si/badge/2026-11-15/8.svg" alt="Ranked #8 on Rubric, 15 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 6 matches source · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Couldn’t checkDate
In November 2025, AppOmni published a demonstration in which a ticket in ServiceNow's Now Assist made an agent recruit a more privileged agent and expose restricted records.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “In November 2025, AppOmni published a demonstration in which a ticket in ServiceNow’s Now Assist made an agent recruit a more privileged agent and expose restricted records”
The source says No passage could be compared.
Source cited by the article: appomni.com/ao-labs/ai-agent-to-agent-discovery-prompt-injection
Matches sourceQuote
Researchers showed that CaMeL's guarantees do not automatically compose between agents, because untrusted data can be reinterpreted as trusted input by a downstream agent.
The cited source says the same thing.
The article says “In October 2026, researchers showed that CaMeL’s guarantees do not automatically compose between agents: “untrusted data can be reinterpreted as trusted input by a downstream agent”
The source says “We find that CaMeL’s guarantees do not compose. We construct a concrete prompt-injection attack that succeeds despite all constituent agents individually operating CaMeL”
Source cited by the article: arxiv.org/abs/2610.05640
Matches sourceQuote
Under RFC 8693, earlier actors in a delegation chain are informational only and are not to be considered in access control decisions.
The cited source says the same thing.
The article says “earlier actors (here, A) are “informational only” and “are not to be considered in access control decisions”
The source says “Prior actors identified by any nested act claims are informational only and are not to be considered in access control decisions”
Source cited by the article: rfc-editor.org/rfc/rfc8693.html
Matches sourceNumber
Nasr et al. (2025) report adaptive attacks against 12 defenses, most with success rates above 90%.
The cited source says the same thing.
The article says “Nasr et al. (2025, preprint) report adaptive attacks against 12 defenses, most with success rates above 90%”
The source says “we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most; importantly, the majority of defenses originally reported near-zero attack success…”
Source cited by the article: arxiv.org/abs/2510.09023
Matches sourceNumber
Yong et al. (2023) found 79% attack success in at least one of four low-resource languages, versus under 1% in English.
The cited source says the same thing.
The article says “79% attack success in at least one of four low-resource languages, versus under 1% in English”
The source says “Nonetheless, combining different low-resource languages increases the jailbreaking success rate to around 79%”
Source cited by the article: arxiv.org/abs/2310.02446
Matches sourceBenchmark result
LatentMAS reports up to 14.6% higher accuracy and 70.8–83.7% fewer output tokens using a shared model.
The cited source says the same thing.
The article says “up to 14.6% higher accuracy and 70.8–83.7% fewer output tokens, using a shared model”
The source says “LatentMAS outperforms advanced single agents and text-based MAS baselines, achieving up to 14.6% higher accuracy, reducing output token usage by 70.8%-83.7%, and providing 4 \times -4.3 \times faster end-to-end inference”
Source cited by the article: arxiv.org/abs/2511.20639
Couldn’t checkDate
Norm Hardy described the confused deputy problem in 1988.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Norm Hardy described this problem in 1988.9”
The source says No passage could be compared.
Source cited by the article: dl.acm.org/doi/10.1145/54289.871709
Matches sourceNumber
Brito and Baquero report that a manifest binding sender, session, message and cache rejected all 295 tested tampered payloads, whereas text-only checks missed cache tampering.
The cited source says the same thing.
The article says “text-only checks missed cache tampering; a manifest binding sender, session, message and cache rejected all 295 tested tampered payloads”
The source says “It accepts 774/774 honest replayed payloads and rejects 295/295 recorded tampered payloads”
Source cited by the article: arxiv.org/abs/2606.28958
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 4 pts, not on front page
Recorded for context only, checked 15 Nov 2026, 20:46 UTC. It played no part in the ranking.