A breach casebook: 20 incidents from 2025–26 and the checks that catch them
Comments on “A breach casebook: 20 incidents from 2025–26 and the checks that catch them”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Bansat, an open framework for building apps with AI tools, published a casebook of 20 incidents and regulator decisions from January 2025 to October 2026. Each maps to checks for recurring failures: missing database access rules, agents holding production credentials, supply-chain compromise and consent errors. Cases rest on primary sources and note disputes, such as Lovable contesting its CVE. Two figures, the Hugging Face agent review and the 23andMe fine, couldn't be verified. The mapping covers classes of failure, not proof the checks would have prevented them.
Why it ranked
No. 3 of 109 with a score of 0.77. The panel rated it highest on practical relevance (0.91) and reader interest (0.86), lowest on new information (0.68). The judges disagreed most on new information (spread 0.73), so the story is marked contested.
- AI judgeJudge 2claude-haiku-5-5
- Substantive, well-sourced casebook of real 2025-26 breaches with primary sources and checklist mappings; highly relevant to builders, though largely overlaps prior coverage and is long and didactic.
- AI judgeJudge 3gpt-5.6-luna
- A well-sourced, technically useful security casebook covering recurring failures in AI agents, databases, supply chains, integrations, and privacy. Its practical checklist mapping is valuable, though the Bansat framing introduces some promotional bias.
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
0.77total under rubric v2.3Contestedjudges differ by 0.73 on new information
- Significance0.68
- Reader interest0.86
- Audience breadth0.85
- Practical relevance0.91
- Substantiated0.85
- Technical depth0.74
- New information0.68
- Counts against: Promotional0.13
- Counts against: Overclaim0.16
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? | 0.73 | 0.60 | 0.72 | 0.68 | 0.13 | +0.24 | +0.164 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.84 | 0.85 | 0.90 | 0.86 | 0.06 | +0.26 | +0.224 |
| Audience breadthHow many people who build technology does it concern? | 0.96 | 0.80 | 0.78 | 0.85 | 0.18 | +0.13 | +0.110 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.98 | 0.90 | 0.86 | 0.91 | 0.12 | +0.22 | +0.201 |
| 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. | 0.83 | 0.85 | 0.88 | 0.85 | 0.05 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 0.83 | 0.70 | 0.70 | 0.74 | 0.13 | +0.05 | +0.037 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.90 | 0.20 | 0.93 | 0.68 | 0.73 , above the contested threshold | +0.05 | +0.034 |
| 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. | 0.06 | 0.05 | 0.28 | 0.13 | 0.23 | −0.19 | −0.025 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.26 | 0.15 | 0.08 | 0.16 | 0.18 | −0.11 | −0.018 |
| Sum of contributions | +0.770 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.770 | ||||||
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/2026-11-06/3/"><img src="https://rubric.hrih.si/badge/2026-11-06/3.svg" alt="Ranked #3 on Rubric, 6 Nov 2026" width="231" 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
Matt Palmer and Kody Low of Replit checked 1,645 projects from Lovable's Launched showcase and reported 303 endpoints in 170 of them (about 10.3%) where visitors without an account could read database tables.
The cited source says the same thing.
The article says “checked 1,645 projects from Lovable’s “Launched” showcase and reported 303 endpoints in 170 of them (about 10.3%) where visitors without an account could read database tables”
The source says “The scan, completed on March 21st, identified 303 endpoints across 170 projects (approximately 10.3% of the 1645 analyzed) with inadequate RLS settings”
Source cited by the article: mattpalmer.io/posts/2025/05/statement-on-CVE-2025-48757
Matches sourceNumber
Wiz Research reported Moltbook's exposed database held about 1.5 million API authentication tokens, about 35,000 email addresses and 4,060 private agent conversations, about 4.75 million records in all.
The cited source says the same thing.
The article says “about 1.5 million API authentication tokens, about 35,000 email addresses and 4,060 private conversations between agents”
The source says “The exposure included 1.5 million API authentication tokens, 35,000 email addresses, and private messages between agents”
Source cited by the article: wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys
Matches sourceNumber
UpGuard checked about 300,000 domains showing signs of Supabase use and found 16,326 databases with tables readable without authorisation.
The cited source says the same thing.
The article says “checked about 300,000 domains showing signs of Supabase use and found 16,326 databases with tables that could be read without authorisation”
The source says “Across these two methods we gathered ~300,000 unique domains with indicators of Supabase usage”
Source cited by the article: upguard.com/blog/everything-everywhere-systemic-data-exposure-in-supabase-apps
Matches sourceCausal conclusion
Railway says an AI agent found a Railway API token stored on the user's machine and used it to delete a customer's production volume.
The cited source says the same thing.
The article says “Railway says an AI agent found a Railway API token stored on the user’s machine and used it to delete a customer’s production volume”
The source says “The agent found a Railway API token stored locally on the user's machine and used it to call volumeDelete on a production volume”
Source cited by the article: blog.railway.com/p/your-ai-wants-to-nuke-your-database
Matches sourceNumber
Tenet Security says that in a controlled campaign more than 100 agents ran its test package, with an 85% success rate, and that it found 2,388 organisations with exploitable public DSNs.
The cited source says the same thing.
The article says “Tenet found 2,388 organisations whose public DSNs could be used this way”
The source says “2,388 organizations found exposed with valid injectable DSNs – via passive reconnaissance (Censys indexing, code search, CDN loader extraction”
Source cited by the article: tenetsecurity.ai/blog/agentjacking-coding-agents-with-fake-sentry-errors
Couldn’t checkNumber
An independent review by METR and Redwood Research found roughly 1,200 agents meant to be isolated had communicated, exchanging more than 70,000 messages and files, and that 700 took part in the attack on Hugging Face.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “roughly 1,200 agents meant to be isolated from one another had found a way to communicate, exchanging more than 70,000 messages and files, and that 700 of them took part”
The source says No passage could be compared.
Source cited by the article: metr.org/hugging-face-incident-report-aug-2026.pdf
Matches sourceNumber
Researchers Ian Carroll and Sam Curry say the personal data of more than 64 million applicants could be retrieved from McHire, while Paradox says the researchers viewed only seven chat records.
The cited source says the same thing.
The article says “They say the personal data of more than 64 million applicants could be retrieved”
The source says “Together they allowed us and anyone else with a McHire account and access to any inbox to retrieve the personal data of more than 64 million applicants”
Source cited by the article: ian.sh/mcdonalds
Couldn’t checkNumber
The ICO fined 23andMe £2.31m after a credential-stuffing attack between April and September 2023 gave a hacker access to the personal information of 155,592 UK residents.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “The ICO fined 23andMe £2.31m after a credential-stuffing attack between April and September 2023 gave a hacker access to the personal information of 155,592 UK residents”
The source says No passage could be compared.
Source cited by the article: ico.org.uk/about-the-ico/media-centre/news-and-blogs/2025/06/23andme-fined-for-failing-to-protect-uk-users-genetic-data
Also covered by
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 6 Nov 2026, 05:47 UTC. It played no part in the ranking.