LLMjacking: The AWS, Google Cloud and Azure settings that cap stolen AI keys
Comments on “LLMjacking: The AWS, Google Cloud and Azure settings that cap stolen AI keys”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Google's Threat Intelligence Group reports that attackers using stolen cloud tokens and AI keys ran up more than $100,000 in a day, and up to $1 million at some organizations, in what it calls LLMjacking. AI News Report turns the advice into vendor settings: per-deployment tokens-per-minute quotas act within minutes, while budget alerts lag, and Bedrock invocation logging and Google Data Access logs are off by default. The figures are GTIG's own, and the LiteLLM CVE it cites could not be verified against public records.
Why it ranked
No. 8 of 137 with a score of 0.75. The panel rated it highest on practical relevance (0.93) and technical depth (0.88), lowest on new information (0.65). Placed No. 8, No. 4 by score alone: the fourth Security story of the day (−0.020), and close to “The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources”, No. 2 on Tuesday 3 November 2026 (−0.004). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Strongly on-topic practical security guidance for cloud AI cost abuse, with specific settings and CVE checks; substantive and actionable, though it builds on a single vendor report and flags unresolved inconsistencies.
- AI judgeJudge 3gpt-5.6-luna
- Highly actionable security guidance for cloud AI abuse, with vendor-specific controls, queries, quotas, and caveats. It substantially elaborates a recently covered incident, so novelty is moderate rather than certain.
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
0.75total under rubric v2.3
- Significance0.66
- Reader interest0.85
- Audience breadth0.72
- Practical relevance0.93
- Substantiated0.77
- Technical depth0.88
- New information0.65
- Counts against: Promotional0.04
- Counts against: Overclaim0.32
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? | 0.72 | 0.55 | 0.72 | 0.66 | 0.17 | +0.24 | +0.159 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.84 | 0.80 | 0.90 | 0.85 | 0.10 | +0.26 | +0.220 |
| Audience breadthHow many people who build technology does it concern? | 0.77 | 0.60 | 0.78 | 0.72 | 0.18 | +0.13 | +0.093 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.94 | 0.90 | 0.94 | 0.93 | 0.04 | +0.22 | +0.204 |
| 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. | 0.70 | 0.75 | 0.86 | 0.77 | 0.16 | +0.05 | +0.039 |
| Technical depthHow technically substantive is it? | 0.99 | 0.85 | 0.78 | 0.88 | 0.21 | +0.05 | +0.044 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.79 | 0.55 | 0.62 | 0.65 | 0.24 | +0.05 | +0.033 |
| 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? | 0.46 | 0.30 | 0.20 | 0.32 | 0.26 | −0.11 | −0.035 |
| Sum of contributions | +0.748 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.748 | ||||||
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/2026-11-04/8/"><img src="https://rubric.hrih.si/badge/2026-11-04/8.svg" alt="Ranked #8 on Rubric, 4 Nov 2026" width="231" height="28"></a>Claim check
Claim check: 7 matches source · 1 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
GTIG reports that in some cases attackers using stolen cloud and AI keys caused over $100,000 in fraudulent charges in one day, and up to $1 million at some organizations before access was revoked.
The cited source says the same thing.
The article says “over $100,000 in fraudulent charges in one day in some cases, and up to $1 million at some organizations before access was revoked”
The source says “we observed fraudulent charges associated with resource consumption exceeding $100,000 USD in a single day, with cumulative charges at some organizations reaching $1 million USD before access was revoked”
Source cited by the article: cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models
Matches sourceNumber
The LLM mixer Jet Night sold an unlimited plan for about $60 a month and, per GTIG, made more than 500,000 calls in under eight hours, passing nearly $1 million in token costs to one victim.
The cited source says the same thing.
The article says “sold an unlimited plan for about $60 a month. GTIG says it made more than 500,000 calls in under eight hours, passing nearly $1 million in token costs to one…”
The source says “Jet Night offers an unlimited token plan for approximately $60 USD per month with alleged access to frontier models”
Source cited by the article: cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models
Matches sourceCausal conclusion
Group UNC6993 used a session-hijacking page to grab access tokens and bypass password and MFA.
The cited source says the same thing.
The article says “One group, UNC6993, used a session-hijacking page to grab access tokens and skip password and MFA”
The source says “which was then posted to the backend API to bypass password and multifactor authentication (MFA) controls”
Source cited by the article: cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models
Matches sourceNumber
AWS says Bedrock model invocation logging, which records request bodies and token counts, is disabled by default and skips the bedrock-mantle endpoint.
The cited source says the same thing.
The article says “Model invocation logging records request bodies and token counts. AWS says it is disabled by default. It also skips the bedrock-mantle endpoint, which has its own quotas”
The source says “Model invocation logging is disabled by default”
Source cited by the article: docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html
Matches sourceNumber
Google's docs say an alerts-only budget does not automatically cap spending, and that spend cap budgets pause a service but enforcement is not instant and overages are billed.
The cited source says the same thing.
The article says “Google's docs say an alerts-only budget "doesn't automatically cap" spending. Its spend cap budgets do pause a service, but enforcement "isn't instant" and overages are billed”
The source says “Aunque es más rápida que los informes, la aplicación de los límites de inversión no es instantánea y los excedentes de costos se facturan con normalidad”
Source cited by the article: docs.cloud.google.com/billing/docs/how-to/budgets-spend-caps
Matches sourceNumber
Google lists Agent Platform predict methods as Data Access (DATA_READ) logs, which are off by default.
The cited source says the same thing.
The article says “Google lists Agent Platform predict methods as Data Access (DATA_READ) logs, which are off by default”
The source says “資料存取 (DATA_READ) 稽核記錄 | datasets.dataItems.list endpoints.explain endpoints.predict endpoints.predictLongRunning endpoints.rawPredict”
Source cited by the article: docs.cloud.google.com/vertex-ai/docs/general/audit-logging
Couldn’t checkNumber
GTIG attributes a LiteLLM /key/regenerate user ID rebind flaw to CVE-2026-40217, but NVD describes that CVE as code execution at /guardrails/test_custom_code.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “NVD describes CVE-2026-40217 as code execution at /guardrails/test_custom_code”
The source says No passage could be compared.
Source cited by the article: nvd.nist.gov/vuln/detail/CVE-2026-40217
Matches sourceNumber
Azure OpenAI quota is assigned as tokens per minute per deployment, and excess requests get HTTP 429.
The cited source says the same thing.
The article says “Azure OpenAI quota is tokens per minute per deployment, and excess requests get HTTP 429”
The source says “When the deployment reaches its TPM limit, further requests receive an HTTP 429 response until token capacity becomes available”
Source cited by the article: learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/quota
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 4 Nov 2026, 05:47 UTC. It played no part in the ranking.