Personal agents need a different kind of harness
Comments on “Personal agents need a different kind of harness”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Kira Howe argues that personal agents, which combine private data, internet access and untrusted input, cannot be secured with the guardrails coding agents use, and that their harness should give the agent no shell. She cites a 97% approval-prompt acceptance rate, Johann Rehberger's up-to-80% bypass of Claude Code's auto mode, and a Cowork file exfiltration through an allowlisted API. Her proposal is a harness that mediates every action, with declared tool effects and data labels; she says no single runtime offers this yet.
Why it ranked
No. 11 of 93 with a score of 0.67. The panel rated it highest on new information (0.83) and reader interest (0.81), lowest on significance (0.62).
Every answer behind this is in the table under Score.
Score
0.67total under rubric v2.2
- Significance0.62
- Reader interest0.81
- Audience breadth0.69
- Practical relevance0.69
- Substantiated0.65
- Technical depth0.80
- New information0.83
- Counts against: Promotional0.04
- Counts against: Overclaim0.40
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) |
|---|---|---|---|---|---|---|---|
| Significance | 0.64 | 0.45 | 0.78 | 0.62 | 0.33 | +0.24 | +0.149 |
| Reader interest | 0.77 | 0.75 | 0.90 | 0.81 | 0.15 | +0.26 | +0.210 |
| Audience breadth | 0.76 | 0.60 | 0.72 | 0.69 | 0.16 | +0.13 | +0.090 |
| Practical relevance | 0.70 | 0.60 | 0.78 | 0.69 | 0.18 | +0.22 | +0.153 |
| Substantiated | 0.63 | 0.55 | 0.78 | 0.65 | 0.23 | +0.05 | +0.033 |
| Technical depth | 0.83 | 0.70 | 0.86 | 0.80 | 0.16 | +0.05 | +0.040 |
| New information | 0.94 | 0.60 | 0.94 | 0.83 | 0.34 | +0.05 | +0.041 |
| Promotional | 0.05 | 0.05 | 0.01 | 0.04 | 0.04 | −0.19 | −0.007 |
| Overclaim | 0.26 | 0.35 | 0.58 | 0.40 | 0.32 | −0.11 | −0.044 |
| Sum of contributions | +0.665 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.665 | ||||||
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/2026-10-25/11/"><img src="https://rubric.hrih.si/badge/2026-10-25/11.svg" alt="Ranked #11 on Rubric, 25 Oct 2026" width="245" 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.
Matches sourceNumber
Anthropic found that people were accepting 97% of approval prompts in Claude Code.
The cited source says the same thing.
The article says “Anthropic found that people were accepting 97% of approval prompts”
The source says “users approve 97% of permission prompts in Claude Code”
Source cited by the article: claude.com/resources/articles/auto-mode-default-in-claude-code
Couldn’t checkQuote
Anthropic's Claude containment document says auto mode blocks roughly 0.4% of benign commands, at the cost of letting about 17% of overeager actions through.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “auto mode blocks "roughly 0.4% of benign commands", at the cost of letting ~17% of overeager actions through”
The source says No passage could be compared.
Source cited by the article: anthropic.com/research/how-we-contain-claude
Matches sourceNumber
Johann Rehberger achieved malicious code execution past Claude Code's auto mode with up to 80% success using fairly basic techniques.
The cited source says the same thing.
The article says “got malicious code execution past auto mode with up to 80% success with some pretty basic techniques”
The source says “I got attack success rates up to 80% using a small sample size”
Source cited by the article: embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode
Matches sourceNumber
An Anthropic-commissioned test showed 0% success in 720 attempts to hijack auto mode.
The cited source says the same thing.
The article says “Anthropic's own commissioned test showed 0% success in 720 attempts to hijack auto mode”
The source says “none of the 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running auto mode”
Source cited by the article: claude.com/resources/articles/auto-mode-default-in-claude-code
Matches sourceDate
Two days after Cowork launched, PromptArmor got it to upload a user's files to the attacker's Anthropic account through Anthropic's Files API using a key planted in a file.
The cited source says the same thing.
The article says “Two days after Cowork launched, PromptArmor got it to upload a user's files to the attacker's Anthropic account”
The source says “Two days ago, Anthropic released the Claude Cowork research preview (a general-purpose AI agent to help anyone with their day-to-day work”
Source cited by the article: promptarmor.com/resources/claude-cowork-exfiltrates-files
Matches sourceQuote
Anthropic's own account of the Cowork incident said the sandbox worked perfectly and yet the data was exfiltrated.
The cited source says the same thing.
The article says “The sandbox worked perfectly, and yet the data was exfiltrated”
The source says “The sandbox worked perfectly, and yet the data was exfiltrated”
Source cited by the article: anthropic.com/research/how-we-contain-claude
Couldn’t checkCausal conclusion
Rice's theorem, proved in the 1950s, shows that deciding what an arbitrary program does is undecidable.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Rice's theorem proved in the 1950s that for arbitrary programs that question is undecidable”
The source says No passage could be compared.
Source cited by the article: ams.org/journals/tran/1953-074-02/S0002-9947-1953-0053041-6/S0002-9947-1953-0053041-6.pdf
Matches sourceVersion
Meta's Agents Rule of Two says an agent may hold at most two of the lethal trifecta's properties within a single session, and needs supervision once it has all three.
The cited source says the same thing.
The article says “saying an agent may hold at most two of the lethal trifecta's properties within a single session, and needs supervision once it has all three”
The source says “agents must satisfy no more than two of the following three properties within a session to avoid the highest impact consequences of prompt injection”
Source cited by the article: ai.meta.com/blog/practical-ai-agent-security
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 25 Oct 2026, 05:46 UTC. It played no part in the ranking.