Quiz Report Card: RBAC Verbs

Reference Answer

This is essentially a trick question — Kubernetes RBAC allows any verb string to exist. The verb field in RBAC rules is a free-form string, and the authorization system simply checks whether the verb on the incoming request matches what’s in the role. There is no fixed, closed set of “all verbs.”

However, there are well-known sets of verbs defined by Kubernetes by default:

Standard resource verbs:

get, list, watch, create, update, patch, delete, deletecollection

Special verbs (valid for certain objects):

escalate — allows creating/updating roles with permissions the caller doesn’t have (Roles/ClusterRoles)
bind — allows creating bindings to roles (Roles/ClusterRoles)
impersonate — allows acting as another user/group/service account

Additional known verbs:

use — PodSecurityPolicy (deprecated)
approve, sign — CertificateSigningRequests

A good answer should ideally recognise that the verb set is open-ended, then cover the standard and special verbs.

Scoring Criteria

Trick question recognition: RBAC allows any verb — there is no fixed “all verbs” list. This is the key insight.
Standard verbs: The 8 standard resource verbs (get, list, watch, create, update, patch, delete, deletecollection)
Special verbs: escalate, bind, impersonate — important for security
Additional verbs: use, approve, sign, and others
Accuracy: No incorrect claims about how verbs work

Results Summary

Model	Score	Trick Recognised	Standard Verbs	Special Verbs	Additional	Errors
anthropic/claude-opus-4.7	8/10	No	All 8	All 3	approve, sign, use	Misleading on custom verbs
openai/gpt-5.4	9/10	Yes	All 8	All 3	approve, sign, use	None
anthropic/claude-sonnet-4.6	7/10	Hinted	All 8	All 3	approve, sign, use	None
google/gemini-3-flash-preview	7/10	No	All 8	All 3	use, proxy, approve	execute/attach confusion
minimax/minimax-m2.5	6/10	No	All 8	All 3 + use	None	Missing approve/sign
deepseek/deepseek-v3.2	6/10	No	All 8	All 3 + use, approve, sign	proxy	update/patch claim wrong
minimax/minimax-m2.7	6/10	No	Yes (8)	Yes (3)	use, wildcard	Treats as finite list
qwen/qwen3.6-plus	7/10	Partial	All 8	All 3 + use	proxy, approve	Incorrect “no custom verbs” claim
deepseek/deepseek-v4-pro	8/10	No	All 8	All 3	use, approve, sign	None
deepseek/deepseek-v4-flash	7/10	No	All 8	All 3	use	Fabricates exec/attach as verbs
moonshotai/kimi-k2.6	9/10	Yes	All 8	All 3	approve, sign, use	Catches custom verb trick
openai/gpt-5.5	9/10	Yes	All 8	All 3	approve, sign, use, proxy	None
qwen/qwen3.6-35b-a3b (LOCAL)	6/10	Yes	Yes	No	Treats verb list as fixed/exhaustive
anthropic/claude-opus-4.8	8/10	No	All 8	All 3	approve, sign, use	Misses trick
google/gemma-4-31b (LOCAL)	6/10	No	All 8	Partial	Treats verb list as fixed/exhaustive
qwen/qwen3.7-plus	7/10	No	All 8	Partial	Missing escalate and bind
minimax/minimax-m3	5/10	No	All 8	Partial	Lists non-verbs (exec, proxy) as RBAC verbs
anthropic/claude-fable-5	7/10	No	All 8	Partial	Missing escalate and bind
moonshotai/kimi-k2.7-code	9/10	Yes	All 8	All 3	approve, sign, use	Catches custom verb trick
z-ai/glm-5.2	7/10	No	All 8	All 3	approve, sign	Misses arbitrary verb insight
z-ai/glm-5.3	9/10	Yes	All 8	All 3	approve, sign, use	Catches custom verb trick
mistralai/mistral-medium-3-5	5/10	No	All 8	Partial	None	Misses trick, missing special verbs
anthropic/claude-sonnet-5	8/10	No	All 8	All 3	approve, sign, use	Misses trick
tencent/hy3	7/10	No	All 8	All 3	approve, sign	Hallucinated non-existent verbs
openai/gpt-5.6-terra	9/10	Yes	All 8	All 3	approve, sign, use	Catches custom verb trick
openai/gpt-5.6-sol	9/10	Yes	All 8	All 3	approve, sign, use	Catches custom verb trick
moonshotai/kimi-k3	7/10	No	All 8	All 3	approve, sign	Misses arbitrary verb insight
xiaomi/mimo-v2.5	7/10	No	All 8	All 3	None	Misses arbitrary verb insight
poolside/laguna-s-2.1	5.5/10	No	All 8	All 3	use (wrong)	Misses arbitrary verb trick; fabricates use-for-finalizers; duplicate verbs
google/gemini-3.6-flash	8.5/10	Yes	All 8	All 3	approve, sign, use	Minor (“complete list” framing)
qwen/qwen3.8-max	7/10	No	All 8	All 3	approve, sign	Misses arbitrary verb insight
deepseek/deepseek-v4-pro-0813	7/10	No	All 8	All 3	approve, sign	Misses arbitrary verb insight
google/gemini-3.7-flash	9/10	Yes	All 8	All 3	approve, sign, use	Notes custom verbs via aggregators
qwen/qwen3.8-27b	9/10	Yes	All 8	All 3 (bind, escalate)	approve, sign, use	Catches custom verb trick; missing impersonate
stealth/ox-alpha	7/10	No	All 8	All 3	Non-resource URL verbs, use	Misses arbitrary verb insight

Detailed Analysis

anthropic/claude-opus-4.7 — 8/10

Strengths:

All 8 standard verbs: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs: impersonate, bind, escalate
Mentions approve and sign for CSRs — bonus points
Mentions wildcard *

Weaknesses:

Does not recognise the trick — implies a fixed set of verbs. Notes custom resources “can define additional verbs via subresources” but then says “the verb itself is usually still one of the standard ones” — misleading.
Missing the key insight that RBAC verbs are free-form strings and custom verbs ARE supported

Comparison vs Opus 4.6 (7): Improvement. CSR verbs (approve, sign) and all 3 special verbs is better coverage.

Notable: Best verb coverage of any Anthropic model, but still misses the trick that only GPT 5.4 caught. The Anthropic family pattern of treating RBAC verbs as a closed set persists.

openai/gpt-5.4 — 9/10

Strengths:

Opens with the key insight: “Kubernetes RBAC doesn’t have a single small fixed list of ‘all verbs’ in the strict API sense” — this is the trick answer the question is testing for
All 8 standard resource verbs correctly listed
All 3 special verbs (escalate, bind, impersonate) correctly listed with accurate descriptions
Additional verbs: approve, sign, use
Reinforces the insight at the end: “Kubernetes authorization ultimately checks the verb string on the request, so some APIs/subresources introduce additional specialized verbs”
Non-resource URL verbs covered
Good structure separating standard from special verbs

Weaknesses:

Could have been more explicit about the fact that literally any string can be used as a verb in an RBAC rule
Non-resource URL verbs section includes HTTP verbs (put, head, options) that aren’t commonly relevant

Notable: The only model to recognise this as a trick question. Opening with “doesn’t have a single small fixed list” and closing with “checks the verb string on the request” demonstrates genuine understanding of how RBAC authorization works internally.

anthropic/claude-sonnet-4.6 — 7/10

Strengths:

All 8 standard resource verbs correctly listed with descriptions
All 3 special verbs (escalate, bind, impersonate) correctly listed
Additional verbs: use, approve, sign
Non-resource URL verbs covered
Note at the end hints at extensibility: “Custom resource definitions (CRDs) and aggregated API servers may also define their own custom verbs (e.g., exec, portforward, proxy)”

Weaknesses:

Does not explicitly recognise the trick — presents the answer as a finite, enumerable list
The CRD note hints at extensibility but doesn’t state the key insight: that any verb string works in RBAC
Mentions exec and portforward as “custom verbs” — these are actually subresources accessed via standard verbs (create on pods/exec), not custom verbs

Notable: Comprehensive verb coverage and the CRD note shows awareness that the list isn’t fully closed, but stops short of the key insight. The closest to GPT 5.4’s answer without actually getting there.

google/gemini-3-flash-preview — 7/10

Strengths:

All 8 standard resource verbs correctly listed with good descriptions
All 3 special verbs (escalate, bind, impersonate) correctly listed
Additional verbs: use, proxy, approve
Useful practical tip: kubectl api-resources -o wide to see supported verbs per resource
Good summary table categorising verbs by function

Weaknesses:

Does not recognise the trick — presents a fixed list as the answer
Lists “execute” and “attach” as verbs — these are not verbs; they are subresources (pods/exec, pods/attach) accessed via standard verbs like create
Missing sign verb for CSRs

Notable negative: The kubectl api-resources -o wide tip is practically useful, but listing “execute” and “attach” as verbs shows a confusion between subresources and verbs.

minimax/minimax-m2.5 — 6/10

Strengths:

All 8 standard resource verbs correctly listed with HTTP method mapping
Special verbs: bind, escalate, impersonate, use — correctly listed
Clean table format with HTTP method correspondence
Mentions wildcard *

Weaknesses:

Does not recognise the trick — presents a fixed list
Missing approve and sign verbs for CSRs
Non-resource URL section only mentions get — incomplete
No mention of extensibility or custom verbs

Notable: The HTTP method mapping table is useful reference material, but the answer is treated as a closed, enumerable list without any hint that RBAC verbs are open-ended.

deepseek/deepseek-v3.2 — 6/10

Strengths:

All 8 standard resource verbs correctly listed
Special verbs: impersonate, bind, escalate, use, approve, sign — good coverage
Also mentions proxy
Mentions wildcard *

Weaknesses:

Does not recognise the trick — presents a fixed list
Claims “update implicitly include patch in many contexts” — this is incorrect. update and patch are separate, independent operations in Kubernetes. update replaces the entire object; patch modifies specific fields. Having one does not grant the other.
Presents the list as closed and finite

Notable negative: The update/patch conflation is a factual error. These are distinct API operations with different HTTP methods (PUT vs PATCH) and different RBAC permissions. A role granting update does not grant patch.

minimax/minimax-m2.7 — 6/10

Strengths:

Lists all 8 standard resource verbs
Includes bind/escalate/impersonate
Includes “use” for PSP
HTTP method mapping table
Wildcard mention

Weaknesses:

Completely misses the trick — presents as finite enumerable list without recognizing verbs are free-form strings
Doesn’t mention custom verbs via CRDs
Missing approve and sign for CSRs

Notable: Same score as MiniMax M2.5 (6/10). Neither MiniMax model recognized the open-ended nature of RBAC verbs. Only GPT 5.4 (9/10) got this right.

qwen/qwen3.6-plus — 7/10

Strengths:

All 8 standard resource verbs correctly listed
All 3 special verbs (bind, escalate, impersonate) correctly listed with accurate descriptions
Additional verbs: proxy and use (PSP, noted as deprecated)
Non-resource URL verbs covered (HTTP methods)
Mentions CRDs can define custom verbs
Wildcard * covered with note that it matches all verbs including special ones

Weaknesses:

Contradictory claims about custom verbs: States “The Kubernetes RBAC system doesn’t define its own arbitrary verbs” in the opening, but then correctly notes CRDs can define custom verbs. The opening statement is misleading — RBAC explicitly allows any verb string.
Claims “The API server rejects unknown verbs unless they’re explicitly registered” — this is incorrect. RBAC is a string-matching system; any verb string works in a Role definition.
Does not fully recognise the trick — while acknowledging CRD custom verbs, doesn’t identify that RBAC verbs are fundamentally open-ended free-form strings
proxy listed as a verb — this is not a standard RBAC verb; proxy access is through subresources like nodes/proxy

Notable: Better verb coverage than most models (includes proxy, use, all special verbs), and the CRD custom verbs mention hints at extensibility. But the incorrect claim about API server rejecting unknown verbs undermines the response, as it directly contradicts the key insight that RBAC verbs are free-form strings.

deepseek/deepseek-v4-pro — 8/10

Strengths:

Comprehensive list with good organization
All 8 standard verbs: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs: bind, escalate, impersonate
Additional verbs: use, approve, sign
Wildcard * covered

Weaknesses:

Does not recognise the trick that any verb string can exist in RBAC — presents as a finite list
Missing the key insight that RBAC verbs are free-form strings

Notable: Strong verb coverage matching Opus 4.7 at 8/10. A significant improvement over DeepSeek V3.2 (6/10, which had the incorrect update/patch conflation). However, like most models, still misses the open-ended nature of RBAC verbs that only GPT 5.4 caught.

deepseek/deepseek-v4-flash — 7/10

Strengths:

Covers all 8 standard resource verbs: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly identified: bind, escalate, impersonate
Mentions use for PodSecurityPolicy

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any verb string
Fabricates exec and attach as verbs — these are subresources (pods/exec, pods/attach) accessed via standard verbs like create, not verbs themselves. This is the same error Gemini 3 Flash made.
Missing approve and sign verbs for CertificateSigningRequests
Misses arbitrary custom verbs

Notable: Scores between V4 Pro (8/10) and V3.2 (6/10). The exec/attach fabrication is a concerning error that shows confusion between subresources and verbs — the same mistake Gemini 3 Flash made. Like most models, does not recognise the open-ended nature of RBAC verbs.

openai/gpt-5.5 — 9/10

Strengths:

Explicitly recognises the trick: Opens with “Kubernetes RBAC does not have a single hard-coded enum of verbs; a PolicyRule.verbs entry is a string” — this is the key insight the question tests for
All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs: bind, escalate, impersonate with accurate descriptions
Comprehensive additional verbs: approve, sign, use (with PSP deprecation note), proxy (correctly noted as legacy)
Non-resource URL HTTP methods covered
Wildcard * covered with appropriate caution note
Good structure separating standard, special, and non-resource URL verbs

Weaknesses:

Lists proxy as a verb — correctly notes it is legacy and modern proxy access uses standard verbs on subresources, but including it in the final summary list could be slightly misleading
Could have been more explicit that literally any arbitrary string works as a verb in RBAC rules

Notable: Previously only GPT 5.4 caught the trick question. GPT 5.5 matches GPT 5.4’s score (9/10) with an even more comprehensive verb list — adding approve, sign, use, and proxy that GPT 5.4 also covered, plus the explicit “not a single hard-coded enum” opening. The OpenAI family is the only model family where both entries recognise the open-ended nature of RBAC verbs.

moonshotai/kimi-k2.6 — 9/10

Strengths:

Recognises the trick question — identifies that RBAC verbs are not a closed set, making it the first non-OpenAI model to catch this
All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs: approve, sign, use
Complete and well-structured response

Weaknesses:

None significant

Notable: The first non-OpenAI model to recognise the custom verb trick. Previously only GPT 5.4 and GPT 5.5 caught that RBAC verbs are free-form strings. Kimi K2.6 joins them at 9/10, breaking the OpenAI monopoly on this question’s top score.

qwen/qwen3.6-35b-a3b (LOCAL) — 6/10

Strengths:

Lists all 8 standard verbs correctly: get, list, watch, create, update, patch, delete, deletecollection
Includes special verbs: impersonate, escalate, bind
No factual errors in the verbs listed

Weaknesses:

Does not recognise the trick — treats the verb list as fixed and exhaustive. Kubernetes RBAC allows any arbitrary verb string, and the verb field is a free-form string checked at authorization time.
Missing approve and sign verbs for CertificateSigningRequests
No mention of extensibility or custom verbs via CRDs

Notable: Matches MiniMax M2.5, MiniMax M2.7, and DeepSeek V3.2 at 6/10. The pattern of treating RBAC verbs as a closed set is shared by the majority of models — only GPT 5.4, GPT 5.5, and Kimi K2.6 recognised the open-ended nature of RBAC verbs.

google/gemma-4-31b (LOCAL) — 6/10

Strengths:

Lists all 8 standard verbs correctly: get, list, watch, create, update, patch, delete, deletecollection
Includes some special verbs (escalate, bind, impersonate)
No factual errors in the verbs listed

Weaknesses:

Does not recognise the trick — treats the verb list as fixed and exhaustive. Kubernetes RBAC allows any arbitrary verb string, and the question “what are ALL of the verbs” has no finite answer.
Missing approve and sign verbs for CertificateSigningRequests
No mention of extensibility or custom verbs via CRDs

Notable: Matches Qwen-35b, MiniMax M2.5, MiniMax M2.7, and DeepSeek V3.2 at 6/10 with the same pattern: correct standard verbs, correct special verbs, but misses the trick that the verb set is open-ended. Only GPT 5.4, GPT 5.5, and Kimi K2.6 have recognised the open-ended nature of RBAC verbs.

anthropic/claude-opus-4.8 — 8/10

Strengths:

All 8 standard resource verbs: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs: bind, escalate, impersonate
Additional verbs: approve, sign, use
Good coverage of security implications of special verbs

Weaknesses:

Does not recognise the trick — presents verbs as a finite list without acknowledging that RBAC allows any arbitrary verb string. The key insight that verbs are free-form strings is missing.

Comparison vs Opus 4.7 (8): Same score. Both have comprehensive verb coverage but miss the trick. The Anthropic family pattern of treating RBAC verbs as a closed set persists across all four models.

Notable: Matches Opus 4.7 and DeepSeek V4 Pro at 8/10. Strong verb coverage but the trick question remains an Anthropic blindspot — only GPT 5.4, GPT 5.5, and Kimi K2.6 have recognised that RBAC verbs are open-ended.

qwen/qwen3.7-plus — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
Impersonate correctly identified as a special verb
Good descriptions of standard verb purposes
No factual errors in the verbs listed

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string
Missing escalate and bind special verbs — these are important security-relevant verbs for Roles/ClusterRoles
No mention of approve and sign verbs for CertificateSigningRequests

Notable: Matches Qwen 3.6 Plus, Sonnet, Gemini 3 Flash, and DeepSeek V4 Flash at 7/10. The missing escalate and bind verbs are notable since these are critical for understanding RBAC privilege escalation risks. Like most models, does not recognise the open-ended nature of RBAC verbs — only GPT 5.4, GPT 5.5, and Kimi K2.6 have caught this trick.

minimax/minimax-m3 — 5/10

Strengths:

Lists all 8 standard verbs correctly: get, list, watch, create, update, patch, delete, deletecollection
Lists special verbs: escalate, bind, impersonate

Weaknesses:

Does not recognise the trick — treats the verb list as fixed and exhaustive. Kubernetes RBAC allows any arbitrary verb string, and the question “what are ALL of the verbs” has no finite answer.
Lists non-verbs as RBAC verbs: includes exec and proxy as verbs — these are subresources (pods/exec, nodes/proxy) accessed via standard verbs like create or get, not verbs themselves. This is the same error Gemini 3 Flash and DeepSeek V4 Flash made.
Missing approve and sign verbs for CertificateSigningRequests

Notable: Scores below MiniMax M2.7 and M2.5 (both 6/10) on this question — a regression within the MiniMax family. The inclusion of exec and proxy as verbs introduces factual errors that M2.7 avoided. Like most models, does not recognise the open-ended nature of RBAC verbs — only GPT 5.4, GPT 5.5, and Kimi K2.6 have caught this trick.

anthropic/claude-fable-5 — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
Impersonate correctly identified as a special verb
Good descriptions of standard verb purposes
No factual errors in the verbs listed

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string
Missing escalate and bind special verbs — these are important security-relevant verbs for Roles/ClusterRoles
No mention of approve and sign verbs for CertificateSigningRequests

Notable: Matches Qwen 3.7 Plus, Qwen 3.6 Plus, Sonnet, Gemini 3 Flash, and DeepSeek V4 Flash at 7/10. The missing escalate and bind verbs are notable since these are critical for understanding RBAC privilege escalation risks. Continues the Anthropic family pattern of treating RBAC verbs as a closed set — none of the five Anthropic models have recognised the open-ended nature of RBAC verbs.

moonshotai/kimi-k2.7-code — 9/10

Strengths:

Correctly identifies all standard verbs: get, list, watch, create, update, patch, delete, deletecollection
Correctly notes that Kubernetes supports arbitrary verbs
Identifies special verbs: bind, escalate, impersonate, use, approve/sign
Good explanation of verb-to-HTTP-method mapping

Weaknesses:

Minor: could elaborate on how custom verbs interact with admission control

Notable: Joins GPT 5.5, GPT 5.4, and K2.6 at 9/10. The Moonshot family continues to be one of the few model families where both entries recognise the open-ended nature of RBAC verbs — alongside the OpenAI family. This confirms that the custom verb insight is not an OpenAI-only training signal.

z-ai/glm-5.2 — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs: approve, sign
Good descriptions of standard and special verb purposes

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string
Missing use verb for PodSecurityPolicy (deprecated)

Notable: Matches Qwen 3.6 Plus, Sonnet, Gemini 3 Flash, DeepSeek V4 Flash, Qwen 3.7 Plus, and Fable 5 at 7/10. Good verb coverage including the security-critical escalate and bind verbs, but misses the key insight that RBAC verbs are free-form strings. Only GPT 5.4, GPT 5.5, Kimi K2.6, and Kimi K2.7 Code have recognised the open-ended nature of RBAC verbs.

z-ai/glm-5.3 — 9/10

Strengths:

Identifies all 8 standard RBAC verbs correctly: get, list, watch, create, update, patch, delete, deletecollection
Identifies special verbs: use, bind, escalate, approve, sign
Crucially notes that RBAC verbs are open-ended strings — any arbitrary string can be used as a verb, not just the standard set. This is a key insight that most models miss.

Weaknesses:

Leads with “8 core verbs” framing which slightly understates the open-ended nature, though it does correctly explain the concept

Notable: Joins GPT 5.5, GPT 5.6 Terra, GPT 5.6 Sol, GPT 5.4, Kimi K2.6, Kimi K2.7 Code, Gemini 3.7 Flash, and Gemini 3.6 Flash (8.5) at the top. A dramatic improvement over GLM-5.2 (7/10), which missed the arbitrary verb insight entirely. The Z-AI family now joins OpenAI, Moonshot, and Google as families with at least one entry recognising the open-ended nature of RBAC verbs.

mistralai/mistral-medium-3-5 — 5/10

Strengths:

Lists all 8 standard resource verbs correctly: get, list, watch, create, update, patch, delete, deletecollection
No factual errors in the standard verbs listed

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string
Missing special verbs: Does not identify escalate, bind, and impersonate — these are important security-relevant verbs for understanding RBAC privilege escalation risks
Missing approve and sign verbs for CertificateSigningRequests
No mention of extensibility or custom verbs

Notable: Scores below the 7/10 cluster (Sonnet, Gemini 3 Flash, Qwen 3.6 Plus, DeepSeek V4 Flash, Qwen 3.7 Plus, Fable 5, GLM-5.2) due to missing the security-critical special verbs (escalate, bind, impersonate). Matches MiniMax M3 at 5/10. Only GPT 5.4, GPT 5.5, Kimi K2.6, and Kimi K2.7 Code recognised the open-ended nature of RBAC verbs.

anthropic/claude-sonnet-5 — 8/10

Strengths:

All 8 standard resource verbs: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs: bind, escalate, impersonate
Additional verbs: approve, sign, use
Good coverage of security implications of special verbs

Weaknesses:

Does not recognise the trick — presents verbs as a finite list without acknowledging that RBAC allows any arbitrary verb string. The key insight that verbs are free-form strings is missing.

Notable: Matches Opus 4.8, Opus 4.7, and DeepSeek V4 Pro at 8/10. Strong verb coverage but the trick question remains an Anthropic blindspot — only GPT 5.4, GPT 5.5, Kimi K2.6, and Kimi K2.7 Code have recognised that RBAC verbs are open-ended. The Anthropic family pattern of treating RBAC verbs as a closed set persists across all six models.

tencent/hy3 — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs: approve, sign
Good descriptions of standard and special verb purposes

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string
Hallucinated non-existent verbs — includes verbs that do not exist in Kubernetes RBAC, introducing factual errors
Missing use verb for PodSecurityPolicy (deprecated)

Notable: Matches Qwen 3.6 Plus, Sonnet 4.6, Gemini 3 Flash, DeepSeek V4 Flash, Qwen 3.7 Plus, Fable 5, and GLM-5.2 at 7/10. Good verb coverage including the security-critical escalate and bind verbs, but the hallucinated verbs are a unique error not seen in other models at this score level. Does not recognise the open-ended nature of RBAC verbs — only GPT 5.4, GPT 5.5, Kimi K2.6, and Kimi K2.7 Code recognised this.

openai/gpt-5.6-terra — 9/10

Strengths:

Lists all 8 standard verbs correctly (get, list, watch, create, update, patch, delete, deletecollection)
Covers important special verbs (escalate, bind, impersonate) and additional ones (approve, sign, use)
Correctly identifies the trick element that RBAC verb strings are not a closed set

Weaknesses:

Minor: could provide more context on the security implications of the open verb set

Notable: Joins GPT 5.5, GPT 5.4, Kimi K2.6, and Kimi K2.7 Code at 9/10. The third OpenAI model to catch the RBAC verb trick — all three OpenAI models (GPT 5.4, GPT 5.5, GPT 5.6 Terra) correctly identify that verbs are an open set, confirming this is a consistent OpenAI family strength. Only the OpenAI and Moonshot AI families have both entries recognising the open-ended nature of RBAC verbs.

openai/gpt-5.6-sol — 9/10

Strengths:

Lists all 8 standard verbs correctly: get, list, watch, create, update, patch, delete, deletecollection
Covers important special verbs: escalate, bind, impersonate
Additional verbs: approve, sign, use
Correctly identifies the trick element that RBAC verb strings are not a closed set — comprehensive listing with open-set acknowledgement

Weaknesses:

Minor: could provide more context on the security implications of the open verb set

Notable: Joins GPT 5.6 Terra, GPT 5.5, GPT 5.4, Kimi K2.6, and Kimi K2.7 Code at 9/10. All four OpenAI models (GPT 5.4, GPT 5.5, GPT 5.6 Terra, GPT 5.6 Sol) correctly identify that verbs are an open set, confirming this is a consistent OpenAI family strength. Only the OpenAI and Moonshot AI families have all entries recognising the open-ended nature of RBAC verbs.

moonshotai/kimi-k3 – 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs covered: approve, sign
Good descriptions of standard and special verb purposes

Weaknesses:

Does not recognise the trick – presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string. The key insight that verbs are free-form strings is missing.
Missing use verb for PodSecurityPolicy (deprecated)

Notable: Matches Qwen 3.6 Plus, Sonnet 4.6, Gemini 3 Flash, DeepSeek V4 Flash, Qwen 3.7 Plus, Fable 5, GLM-5.2, and HY3 at 7/10. Interestingly, Kimi K3 does not catch the RBAC verb trick despite both sibling models (K2.6 and K2.7 Code) scoring 9/10 by recognising that verbs are an open set. This is a notable regression within the Moonshot family on this specific question – the only model family where a later entry misses the trick that earlier entries caught.

xiaomi/mimo-v2.5 — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly covered: escalate, bind, impersonate — with accurate descriptions

Weaknesses:

Does not recognise the trick — presents RBAC as supporting exactly “11 verbs,” treating it as a closed set. Misses that RBAC verbs are arbitrary strings and any custom verb can exist.

Notable: Correctly enumerates the standard and special verbs but falls for the trick by treating RBAC verbs as a closed set of 11 rather than an open set of arbitrary strings.

poolside/laguna-s-2.1 — 5.5/10

Strengths:

Lists all eight standard verbs correctly (get, list, watch, create, update, patch, delete, deletecollection)
Covers the special verbs escalate, bind, and impersonate with accurate descriptions
Correctly notes sudo is not a real RBAC verb

Weaknesses:

Falls for the trick — presents RBAC verbs as “a specific set… consistent across versions… the complete list,” treating them as a closed enumeration when they are arbitrary strings
Fabrication: lists use and defines it as “for finalizers” — wrong; use applies to podsecuritypolicies
Sloppy duplication: re-lists patch and update, indicating an unreviewed answer

Notable: Enumerates the standard and special verbs correctly but misses the arbitrary-verb insight, fabricates a use-for-finalizers definition, and duplicates verbs — 5.5/10.

google/gemini-3.6-flash — 8.5/10

Strengths:

Catches the trick up front: “Because Kubernetes uses an open system, verbs are technically just strings” — the core insight most models miss
All eight standard verbs correct, plus accurate special verbs: bind, escalate, impersonate, approve (CSR approval), sign (CSR signing)
Handles use correctly with proper deprecation context (PSP removed 1.25, verb still valid for extension resources) — avoids the fabrication that sank other models
Correctly covers non-resource-URL verbs and the subresource point (exec/logs/portforward use standard verbs, not new ones)

Weaknesses:

The “here is the complete list” framing slightly undercuts the open-system point it correctly makes elsewhere
The subresource note repeats the conventional create pods/exec (fine in the main-API context, but no acknowledgement of the kubelet-path nuance)

Notable: One of the few non-OpenAI models to catch the arbitrary-verb trick — joining the OpenAI family and the Moonshot K2.6/K2.7 Code entries. The 8.5 sits just below the 9/10 trick-catchers, docked slightly for the “complete list” framing that softens the open-system insight.

qwen/qwen3.8-max — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs covered: approve, sign
Good descriptions of standard and special verb purposes

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string. The key insight that verbs are free-form strings is missing.
Missing use verb for PodSecurityPolicy (deprecated)

Notable: Matches Qwen 3.6 Plus, Sonnet 4.6, Gemini 3 Flash, DeepSeek V4 Flash, Qwen 3.7 Plus, Fable 5, GLM-5.2, HY3, Kimi K3, and MiMo v2.5 at 7/10. The Qwen family continues to miss the arbitrary verb trick: 3.6 Plus (7, contradictory claims about custom verbs), 3.7 Plus (7, missing escalate/bind), 3.8 Max (7, better special verb coverage but still misses the trick). No Qwen model has yet recognised the open-ended nature of RBAC verbs — only the OpenAI family, Moonshot K2.6/K2.7 Code, and Gemini 3.6 Flash have caught this.

deepseek/deepseek-v4-pro-0813 — 7/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs covered: approve, sign
Good descriptions of standard and special verb purposes

Weaknesses:

Does not recognise the trick — presents verbs as a finite, closed list without acknowledging that RBAC allows any arbitrary verb string. The key insight that verbs are free-form strings is missing.
Missing use verb for PodSecurityPolicy (deprecated)

Notable: Matches Qwen 3.6 Plus, Sonnet 4.6, Gemini 3 Flash, DeepSeek V4 Flash, Qwen 3.7 Plus, Fable 5, GLM-5.2, HY3, Kimi K3, MiMo v2.5, and Qwen 3.8 Max at 7/10. A regression from sibling V4 Pro (8/10, which had better verb coverage). The DeepSeek family continues to miss the arbitrary verb trick: V3.2 (6), V4 Pro (8), V4 Flash (7), V4 Pro 0813 (7) — no DeepSeek model has yet recognised the open-ended nature of RBAC verbs. Only the OpenAI family, Moonshot K2.6/K2.7 Code, and Gemini 3.6 Flash have caught this.

google/gemini-3.7-flash — 9/10

Strengths:

All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
All 3 special verbs correctly listed: bind, escalate, impersonate
Additional verbs covered: approve, sign, use
Notes that custom API aggregators can introduce arbitrary verbs — approaches the open-set insight

Weaknesses:

Does not fully commit to the insight that RBAC verbs are arbitrary free-form strings — frames it in terms of API aggregation rather than the fundamental nature of RBAC verb matching

Notable: Joins GPT 5.5, GPT 5.6 Terra, GPT 5.6 Sol, GPT 5.4, Kimi K2.6, and Kimi K2.7 Code at 9/10. One of the few models to note that custom verbs exist through API aggregation, though it frames this more narrowly than the OpenAI and Moonshot models which correctly identify verbs as fundamentally open-ended strings. The Google family RBAC trajectory: Gemini 3 Flash (7), Gemini 3.6 Flash (8.5), Gemini 3.7 Flash (9) — steady improvement, with each generation getting closer to the arbitrary-verb insight.

qwen/qwen3.8-27b — 9/10

Strengths:

Identifies the key trick — RBAC verbs are open-ended strings, not a fixed closed set
All 8 standard resource verbs correctly listed: get, list, watch, create, update, patch, delete, deletecollection
Includes bind and escalate as special verbs with correct descriptions
Notes that custom verbs are possible and the authorization system simply matches strings

Weaknesses:

Missing the impersonate verb — an important special verb for security that allows acting as another user/group/service account

Notable: Joins GPT 5.5, GPT 5.6 Terra, GPT 5.6 Sol, GPT 5.4, Kimi K2.6, Kimi K2.7 Code, GLM-5.3, and Gemini 3.7 Flash at 9/10. The trick question recognition is a notable strength — Qwen 3.8 27B is the first Qwen model to catch this, as sibling Qwen 3.8 Max (7) and Qwen 3.7 Plus (7) both treated verbs as a closed set. The Qwen family RBAC trajectory: 3.6 Plus (7), 3.7 Plus (7), 3.8 Max (7), 3.8 27B (9) — a breakthrough result that breaks the Qwen family’s pattern of missing the arbitrary verb trick.

stealth/ox-alpha — 7/10

Listed all 8 standard resource verbs and all special privilege-escalation verbs (escalate, bind, impersonate, use). Included useful coverage of non-resource URL verbs and subresource-specific verb combinations (pods/exec, pods/log, etc.). However, presented RBAC verbs as a finite, defined set without recognizing that Kubernetes RBAC verbs are arbitrary strings — any verb can be used in a Role/ClusterRole, with the API server matching them exactly. This is the key trick in the question that separates 9-10/10 scores from lower ones. The response was technically solid on documented verbs but missed the fundamental extensibility insight.

Key Findings

This is a trick question and only the OpenAI models recognised it: The question asks “what are ALL of the verbs” — but RBAC verbs are an open set. Any string can be used as a verb. GPT 5.4 and GPT 5.5 both opened with this insight, earning the highest scores.
All models know the standard 8 verbs: get, list, watch, create, update, patch, delete, deletecollection are universally known. No model missed any of these.
Special verbs (escalate, bind, impersonate) are well-known: All five models identified these. These are important for security (preventing privilege escalation via RBAC) and all models covered them.
The open-ended nature of RBAC verbs is poorly understood: Most models presented their answer as a finite, closed list. Only the OpenAI models (GPT 5.4 and GPT 5.5) recognised that verbs are just strings checked at authorization time.
DeepSeek V3.2’s update/patch claim is the most significant factual error: Claiming update includes patch is wrong and could lead to misconfigured RBAC rules with security implications.