Pluto cross-cutting reconciliation — 2026-11-12
This is a bounded reconciliation of Pluto’s developer-tool, CI and cross-cutting research, not a full currency audit of every mentioned vendor. Source articles were read from the complete Pluto sitemap crawl in this run; underlying capability changes below were separately checked against first-party documentation. No tenant or live policy was changed. Historical exploit reproduction and malware samples were not executed.
Material changes and primary verification
| Finding | Source and primary verification | Disposition |
|---|---|---|
| Imports can preserve an ID while changing publisher/package; popularity is not authenticity | Pluto importer; Cursor extensions, fetched 2026-11-12, explicitly confirms cross-registry differences | Cursor8.1 removes branding/download-count trust ratings and requires actual registry/publisher/package review; inventory pack stops presenting counts as vetting. Devin Desktop2.1/2.2 already enforces IDs/gallery; adds destination-registry approval and import testing. No universal current silent-import claim or guessed auto-update policy. |
| A valid signature does not prove an authorized build | Pluto CI lessons; GitHub CLI verifier, fetched 2026-11-12 | GitHub3.5 corrected; actual --signer-workflow, --source-ref, --source-digest contract enforced in CLI3.20. Reusable signer distinguished from caller. Removes placeholder artifact and mutable OCI latest invocation, propagates failure. |
| Executable agent/IDE/MCP configurations need code-owner review | Same Pluto CI source; GitHub CODEOWNERS, fetched 2026-11-12 | GitHub3.11 extends paths; config3.30 emits mergeable syntax instead of destroying existing CODEOWNERS. Protects CODEOWNERS itself; documents last-match precedence, owner access, branch/ruleset enforcement and noncoverage of local unreviewed checkout. |
| Legacy CircleCI MCP patch advice is superseded by vendor deprecation | Pluto CircleCI research; official repository and current MCP overview, fetched 2026-11-12 | New CircleCI1.4: remove legacy npm connection, choose hosted/CLI intentionally, avoid representing hosted as read-only, test approvals. CLI1.04 inventories documented circleci help reference without starting MCP or reading credentials. |
| Community GitLab MCP is a separate credential-bearing product | Pluto GitLab research; maintainer file-read advisory and SSRF advisory, fetched 2026-11-12 | GitLab1.3 maps incident to existing token inventory/scope/expiry control. No invented GitLab toggle for third-party server. Advisory patched2.1.27 is historical floor, not current blanket assurance. |
Automation census for changed capabilities
All six surfaces were considered; this is capability-scoped, not a claim to have re-enumerated every API of each platform. Existing unrelated packs remain unchanged.
| Surface | GitHub provenance / CODEOWNERS | Extension provenance (Cursor/Devin) | CircleCI MCP |
|---|---|---|---|
| Terraform | Existing ruleset/protection packs; no new verifier implementation warranted | No verified cross-registry attestation/enforcement resource in reviewed references | No verified MCP migration resource in reviewed references |
| API | GitHub REST can retrieve attestations; does not by retrieval alone enforce verifier policy | No verified API proving installed package publisher identity; do not confuse inventory with authenticity | Hosted MCP documented; no tenant-wide migration/approval API inferred |
| CLI | gh attestation verify current flags verified; move old misfiled api shell to cli |
Existing inventory only; no authenticity claim | Official circleci help reference current documented inventory; does not enforce MCP authorization |
| SDK | No new SDK functionality used; CLI supplies signature/identity verifier | No relevant verified admin SDK introduced | No new SDK functionality used |
| Config | Native CODEOWNERS generator; existing branch protection required | Existing allowlist/gallery and cooldown packs retained, approval guidance clarified | MCP host configuration varies by client, not a universal CircleCI setting |
| SIEM | Existing audit coverage; no invented provenance event schema | No invented extension provenance event | Existing CircleCI audit controls retained; no invented MCP event or complete-tool-coverage claim |
CLI inventory rows consulted: GitHub38, CircleCI19 and current Cursor/Cognition rows. New CLI packs are first-party read/verify operations. No token acquisition, mutation, dependency installation or malware execution occurs in tests. scripts/test-pluto-crosscut-packs.py passes nine offline tests covering identity flags, failure propagation, bad source/digest/profile rejection, option-delimited artifact paths, non-destructive CODEOWNERS generation, owner injection rejection and CircleCI help-only invocation. These tests do not validate live signature or policy enforcement.
Source-by-source broader dispositions
| Pluto source | Reconciliation / exclusion |
|---|---|
| TeamPCP campaign and full story | Incident chronology and defender context. Actionable seven-lessons recommendations reconciled below; no campaign scanner/code executed or unsigned IoC feed treated as enforcement. |
| Seven CI lessons | Build/release separation and least-privilege tokens: GitHub3.2/3.8/5.1/5.2; immutable references:3.10/6.2; provenance:3.5 corrected; IDE/config review:3.11 expanded; release age:6.7 already covers native Dependabot cooldown and Renovate; deterministic dependency controls:6.1/6.2/6.8; response:6.6. Generic host isolation/EDR detections and incident credential rotation remain operations/howtorotate scope, not a new SaaS configuration API. |
| Importer syndrome | Cursor and Devin changes above. No Kiro guide exists; research supplies an import-review pattern, not enough first-party evidence for a fabricated Kiro policy pack. Publisher-side registry anti-fraud changes are not customer-controllable settings. |
| Cursor security issues | Existing Cursor4.1/4.2/4.3 MCP restrictions,5.1/5.2 approvals,6.1/6.2 rules review,7.1 workspace trust,10.2 telemetry,11.3 hooks. High-level education/threat modeling is not additional native controls. |
| Cursor AI coding | Same code review, endpoint monitoring and agent-boundary themes; existing Cursor controls above, no new verified native setting. |
| Top MCP risks | Inventory, approved servers, least privilege, tool confirmations and egress map Cursor4.1–4.3/5.2 and product-specific peer guides. Token-audience validation and HTTP endpoint authentication are MCP server implementation responsibilities, not assumed customer-configurable switches. Do not transplant blanket ‘every action logged’ claims into platform coverage. |
| MCP debrief | Artifact/startup-command provenance, privileges and transport inventory map Cursor4.1 and review controls; source6.1/6.2 and GitHub3.11 now cover config drift review. Version pinning alone does not authenticate a remote mutable server. Generic service binding/SSRF implementation is outside SaaS pack scope. |
| Wide open MCPs | Exposure research supports inventory and integration permission review; generic internet server deployment scans/patching are not vendor tenant controls and were not run. |
| Clean scans | A clean static/metadata scan is not a safety guarantee or complete runtime assessment. Existing Cursor4.1 approval/provenance and4.3 egress remain required; no scanner adoption or clean-result allow rule introduced. |
| Supply-chain risks / skills | Static review, dependency provenance, sandbox and runtime checks map Cursor6.1/6.2/8.3 and Claude-specific reconciliation. Generic cryptographic provenance is not proof that every skills marketplace offers signatures. |
| Agent supply-chain attacks | Broad third-party code/dependency governance; GitHub3.11/6.1/6.2/6.7 and Cursor4.1/8.3 cover implementable portions. No unique fetched product-admin mechanism. |
| AI supply-chain risk | Same asset inventory/provenance/least-privilege pattern; no additional verified vendor setting. |
| Reused models/prompts/apps | Model registry integrity and custom AI application design are generic engineering/infrastructure; SaaS integration principles map existing token/repository/agent controls. No new model execution guide. |
| Agent governance | Human/agent attribution, scoped tools, repository boundaries, evidence and revocation map product-specific identity, approvals and audit controls. Per-agent distinct identity/JIT support is not assumed universally. |
| Agent authorization | Conceptual architecture, not evidence every product supports per-call JIT or dedicated identities. Apply existing provider-account permissions; no fictitious IAM protocol/support claim. |
| CircleCI MCP | Current migration/boundary gap implemented above, not the obsolete npm server’s historical environment-variable baseline. |
| GitLab MCP | Integration token context added above; no claim community server is GitLab-owned or fixed by GitLab tenant patch. |
| Apify prompt injection | No Apify guide exists. Separate untrusted retrieval from privileged mutation using client-level approvals/scopes/egress (Cursor4.1–4.3/5.2); no invented Apify setting to eliminate prompt injection. Article’s historical SSRF fixes are implementation patch advisories, not a complete tenant baseline. |
| Jupyter notebook backdoor | Self-hosted server route authentication and DNS-rebinding patch, no Jupyter SaaS tenant control demonstrated. Excluded from new code pack; relevant client server inventory remains covered. |
| Total Recall | Self-hosted memory server vulnerabilities and service patching, not authority for a ChatGPT/Claude memory toggle. Product memory/retention controls reconciled independently by owning workers. |
| MCPwnfluence | Community mcp-atlassian implementation path/SSRF fixes; not Atlassian’s official hosted MCP. No propagation of community flags into Jira/Confluence admin APIs. General reviewed-server/scoped-token pattern covered. |
| Nginx MCP | Self-hosted nginx-ui service authentication/patching is infrastructure scope; no new SaaS guide. |
| MCP exploitation risk | Research follow-up to MCPwnfluence; same community implementation disposition, no independent tenant control. |
| MCP security integration | Crosscut protocol risks covered through product MCP policy and upstream scopes; not a source for unsupported server-auth toggles. |
| Hugging Face config injection | Transformers model-loading/library execution mitigation, not a SaaS tenant setting; exclude generic model-runtime/infrastructure pack. |
| Load/Mine | Malware/model artifact investigation; no malware downloaded/executed. Package provenance/least privilege are transferable but no new HTH Hub control asserted from incident alone. |
| Count Dooku, SleepyDuck, PhantomBoard, Below Deck, Nebula Junior, Another supply chain | Extension/campaign research reinforces Cursor8.1/8.2 and Devin2.1/2.2; quarantine/host recovery and IoC scanning remain incident response. No download-count heuristics or claims that allowlisting alone cures publisher compromise. |
Additional general-context articles: final dispositions
All 26 cached article bodies were reviewed on 2026-11-12. These dispositions distinguish transferable security principles from evidence of a customer-configurable product mechanism. None supplied an additional verified native control gap beyond the product-specific work above and peer ledgers.
| Pluto source | Final disposition |
|---|---|
| vibe-coding-security | General secure-development guidance: mandatory review, SAST/SCA/secrets checks, accountable owners and inventory. Maps to GitHub repository review/CODEOWNERS (3.11) and existing security-scanning/dependency controls, plus Cursor privacy and approvals. Training, artifact tagging and architecture review are process requirements, not an additional vendor admin switch; no new pack. |
| the-rise-of-shadow-ai-how-ai-built-apps-escape-security-visibility | Discovery/governance context: CASB/network inventory, risk-tiering, procurement review, approved alternatives and ownership. Existing product identity, approved integrations and audit controls implement applicable portions. No named enforceable product setting/API; do not generalize the article’s continuous-learning claim to every hosted model. |
| next-gen-endpoint-protection-vs-ai-risks | EDR selection and architecture commentary: pilot/tune policies, inventory local AI, correlate identity and endpoint evidence. No concrete EDR policy contract or tenant setting supplied. Excluded as product-neutral operational guidance; no universal cross-OS parity, zero-day prevention or unmanaged-device enforcement guarantee adopted. |
| citizen-development-security-teams | Asset ownership, connector scoping, access review, native logging and development guardrails. Maps to product-specific connector authorization/audit controls and the separately reconciled Copilot Studio baseline. No platform-native setting beyond that evidence; generic organizational training/risk classification excluded from executable packs. |
| enterprise-ai-security-at-scale | Inventory agents/prompts/models, review generated artifacts, scoped identities, risk reviews and monitoring. GitHub3.11 now covers executable agent configuration review; Cursor4.1–4.3 and5.2 cover reviewed tools/egress/approvals. Dedicated identities, JIT and complete action logs are desired architecture, not universally supported SaaS features; no new generic API. |
| ai-application-security-risk-fast-development | General AppSec/SDLC guidance: scanning, dependency and secret checks, secure coding, shadow-AI visibility and compliance mapping. Existing GitHub scanning and dependency controls plus agent permissions cover concrete HTH portions. Does not identify a new customer-controlled vendor capability; no claim an AI scanner guarantees secure code. |
| ai-offensive-cyber-operations | Offensive-capability research and threat-model commentary, not a hardening implementation guide. Faster discovery, continuous exposure review and realistic testing reinforce existing controls; no new native setting or pack. Study success rates are scenario-specific and were not promoted to control-effectiveness guarantees; no exploitation performed. |
| ai-code-security-defenders | AI-assisted review/scanning advocacy and defensive-program context. Repository review, CI checks and dependency scanning map existing GitHub controls. No concrete product control contract; excludes unverified universal lower-false-positive/faster-detection claims and does not replace conventional checks or human validation with model output. |
| building-mythos-security | Security-program synthesis: dynamic inventory, patch validation, least privilege, MFA, egress and reviewed model access. Maps to existing identity, GitHub dependency/build controls and Cursor4.3/2.2. VulnOps organization, autonomous patching and response-time prescriptions are operational program choices, not new SaaS hardening code; no automatic production containment introduced. |
| continuous-monitoring-security | Telemetry architecture across prompts, tools, builds, identities and runtime. Existing audit/SIEM capabilities remain product-scoped; GitHub provenance and configuration-review corrections address pipeline integrity. No verified new event/API schema; does not assume access to hidden chain-of-thought or complete prompt/tool logs, or that monitoring prevents all injection. |
| vibe-coding-meaning-origins-use-cases | Educational definition/history/use cases with code review, testing and governance requirements. Maps existing development/repository controls; no distinct product hardening mechanism. Product examples do not establish that every platform offers the same RBAC, audit or compliance features. |
| real-value-of-agentless-monitoring | Monitoring architecture and buyer guidance: pilot scope, minimum read permissions, SIEM integration, tuning and coverage review. No specific vendor module/API supplied; infrastructure-focused examples excluded from new SaaS packs. Explicitly do not adopt marketing claims of 100% visibility, zero overhead or guaranteed capture of every ephemeral resource. |
| enterprise-ai-guardrails | Policy-to-enforcement architecture: approved models/integrations, data classification, least privilege, approvals and audit. Maps Cursor2.2 and4.1–4.3/5.2 plus product-specific worker controls. Universal prompt interception, per-agent identity/JIT and auto-revocation are not inferred from this article; no new verified native gap. |
| endpoint-visibility-agentic-era | Layered endpoint/application/identity telemetry guidance. Existing agent audit controls and monitoring notes cover supported surfaces; no new vendor telemetry API. Retain EDR as complementary, not ineffective; desired decision-chain visibility is not a documented capability guarantee. |
| gartner-report-for-ai-endpoint-security | Analyst-report recap and Pluto promotion; inventory/application-control/runtime themes map existing reviewed-extension and least-privilege controls. No new configuration interface. Analyst recognition is not endorsement or technical proof; no new pack and no claim to have read gated Gartner reports. |
| vs-code-security-risk | Extension execution, workspace trust and remote-session boundary risks map Cursor7.1/8.1/8.2 and Devin Desktop2.1/2.2; the importer corrections strengthen provenance. Generic extension CVEs and host recovery are maintenance/incident-response scope, not new SaaS APIs. Do not infer all remote hosts are automatically compromised or sandboxed. |
| vibe-coding-safe | Secure prompts are not enforcement: require review, scanning, dependency validation and threat modeling. Existing GitHub checks and Cursor approvals/rules boundaries cover implementable portions. Benchmark commentary does not introduce a vendor setting or prove generated code safe; no new pack. |
| introducing-plutonium | Resource-hub/marketplace announcement with useful add-on admission criteria: scoped privileges, inspectable pinned packages, no silent settings changes, approval gates, egress ownership and untrusted-output handling. Maps Cursor4.1–4.3/5.2/6.1–6.2/8.3 and GitHub3.11. Catalogs/hands-on guides handled by ecosystem workers; marketplace inclusion is not an unconditional safety guarantee or automatic allowlist. |
| ai-agents-cybersecurity | ARTEMIS penetration-testing study commentary; false-positive validation, contextual limits and human oversight. Excluded from new product controls: no customer-admin mechanism, no experiments executed, and no study result generalized to a universal attack/defense benchmark. |
| agentic-ai-security | Threat-model taxonomy: ambiguous goals, excessive tools, multi-agent feedback, memory poisoning and untrusted inputs. Maps existing agent permissions, approvals and egress controls (Cursor4/5); no new native API. Do not adopt blanket claims that IAM/network segmentation are bypassed or hidden reasoning is available for logging. |
| mcp-inspector | Pluto scanner launch plus CircleCI/Jupyter/Apify incident recaps. CircleCI migration implemented above; other incidents already separately disposed. Optional package-analysis product, not a new vendor hardening setting; no code/packages uploaded or scanner installed, and a clean static verdict is not runtime assurance. |
| latio-2026-ai-security-market-report | Market report recap and vendor recognition; emphasizes control-point coverage tradeoffs. Context-only exclusion from packs: no configuration mechanism and no technical assurance inferred from market placement. No claim to have independently evaluated every cited product/report. |
| welcome-to-planet-pluto | Funding/company-launch and mission announcement. No product-admin hardening setting, executable configuration or new control requirement; excluded as marketing/company context. |
| prompt-injection-attacks | Concrete layered design principles, not a new native switch: untrusted retrieval must not authorize actions, minimize tool/data access, approve consequential actions, constrain egress and test boundaries. Maps Cursor4.1–4.3/5.2 and product-specific agent hardening. No prompt/filter/RAG guarantee or complete hidden-reasoning logging claim; no additional pack warranted. |
| nist-ai-rmf | Voluntary governance framework interpretation: GOVERN/MAP/MEASURE/MANAGE, inventory, measured evaluations and incident integration. Context-only framework mapping, not new SaaS configuration. The illustrative 0.5% threshold is not a NIST requirement or HTH acceptance baseline; no certification/compliance claim added. |
| zero-trust-ai-agents | Action-level authorization, scoped delegation, short-lived authority and safe workflow termination architecture. Existing token/approval/egress controls implement supported portions. Entra reference is not proof all vendors support agent-bound identity/JIT/transaction rollback; no invented universal authorization interface or extra pack. |
Limits and validation
GitHub guide maturity was narrowed to ai-drafted after the substantive provenance/CODEOWNERS corrections; the changed3.11 ClickOps badge was removed. Historical marks on untouched surfaces describe the prior dated run, not this offline pass. Baseline pack validator traverses the full corpus despite vendor argument and reported pre-existing failures outside this work; final touched validation and generated YAML sync belong to the coordinating pass. No new ai-validated or ni-* badge is claimed. No full vendor Trust Center refresh is claimed for this focused source-led correction; source admission and concrete capability documents, not assurance marketing, ground the changes.