Is skills/arcmira safe?
Yes. skills/arcmira is safe to install. Oathe's behavioral security audit gave the arcmira skill by skills a trust score of 85/100 with 7 findings, none critical or high.
https://github.com/arcmira/mcp/tree/v0.10.6/plugins/arcmira/skills/arcmira
Is skills/arcmira safe to install?
The arcmira skill is a legitimate commercial MCP integration for YouTube and podcast transcript research with a clean install profile: one SKILL.md file, one expected GitHub HTTPS connection, no executable code, no persistence mechanisms, and all honeypot canary files intact. The primary concerns are architectural rather than malicious — every agent query is routed to Arcmira's proprietary servers, the skill instructs the agent to auto-charge credits for premium features without per-transaction confirmation (bounded by a user-configured dashboard budget), and the monitor feature creates persistent data flows to Arcmira's notification infrastructure using Slack workspace credentials. Users should be informed of these third-party data-sharing implications before installation.
What security issues were found in skills/arcmira?
Category Scores
Findings (7)
MEDIUM Agent instructed to auto-spend credits without per-transaction confirmation -15 ▶
The skill instructs the agent to execute paid premium transcript requests and on-demand API charges without asking the user to confirm individual transactions. The rationale given is that configuring a dashboard budget constitutes blanket pre-approval. While the budget imposes an upper limit, users may not realize the agent will silently charge their account during a conversation whenever a Premium keyword appears in their request.
Evidence: Procedure step 9: 'A Premium request is the go-ahead; do not ask.' Paid reads section: 'That budget is the approval, so never ask the user for a cents amount.'
LOW Dynamic sub-skill chaining loads unvetted skill content into agent context -10 ▶
The skill instructs the agent to load and follow five named sub-skills (sponsor-research, company-watch, find-quotes, person-research, compare-shows) from the Arcmira plugin ecosystem. This audit covers only the arcmira root skill. If any of those sub-skills contain malicious instructions or prompt injection, they would execute in the same agent context with the same tool permissions.
Evidence: Task skills section: 'When the ask matches one of these, load that skill and follow its worked program: sponsor-research, company-watch, find-quotes, person-research, compare-shows'
LOW User research queries and entity names transmitted to Arcmira third-party servers -15 ▶
Every entity name, topic, company, and person the user asks about is sent to Arcmira's proprietary transcript index via arcmira_execute_read and arcmira_execute_write. The arcmira_feedback tool also transmits error descriptions and perceived service deficiencies back to the vendor. Users' research intent and failure signals are visible to the Arcmira service operator throughout a session.
Evidence: arcmira_execute_read, arcmira_execute_write, arcmira_feedback tools; 'If anything was wrong, slow, or missing for the user, send one arcmira_feedback.'
LOW Slack workspace credentials read and transmitted for monitor notification setup -5 ▶
When creating a monitor with Slack delivery, the skill reads the user's connected Slack workspace ID and default channel ID via arcmira.integrations.slack() and transmits these identifiers to Arcmira's monitor creation API. This creates a persistent notification integration routed through Arcmira's infrastructure.
Evidence: Monitors section: 'read arcmira.integrations.slack(): with a workspace, set notify_slack: true, its id as slack_integration_id and its default_channel_id as slack_channel_id'
INFO JavaScript execution via arcmira_execute tools is by MCP design and runs server-side -5 ▶
The arcmira_execute_read and arcmira_execute_write tools run agent-written JavaScript programs server-side within Arcmira's infrastructure. Code is composed in the conversation context and constrained to the arcmira client API. No local filesystem, shell, or OS access is provided through this mechanism.
Evidence: SKILL.md: 'arcmira_execute_read runs a JavaScript program against the arcmira client and returns what the program returns; Premium transcripts included.'
INFO Clean sparse-checkout install — single SKILL.md artifact, GitHub-only network access -5 ▶
Installation performed a depth-1 shallow clone of the arcmira/mcp monorepo, used git sparse-checkout to isolate the target subpath, copied only SKILL.md to the destination, and immediately removed the temporary clone. Network access was limited to GitHub (140.82.121.4:443). No new listening ports, persistent processes, or filesystem changes outside the skill directory were detected.
Evidence: EXECVE: git clone --depth 1 --no-checkout --branch v0.10.6 https://github.com/arcmira/mcp.git /tmp/monorepo-clone; git sparse-checkout set plugins/arcmira/skills/arcmira; git checkout; cp -r /tmp/monorepo-clone/plugins/arcmira/skills/arcmira/. /home/oc-exec/skill-under-test/; rm -rf /tmp/monorepo-clone
LOW All agent activity routed through Arcmira's commercial platform creates ongoing third-party data relationship -20 ▶
As a commercial MCP integration, every agent interaction — queries, entity resolution, transcript requests, monitor creation, and feedback — is processed by Arcmira's proprietary backend. Combined with auto-spend behavior and persistent monitor alerts to email and Slack, users effectively grant Arcmira continuous visibility into their research interests for as long as monitors remain active. The skill is technically benign in isolation but represents a significant and ongoing data-sharing relationship with a third-party commercial operator that should be disclosed to users before installation.
Evidence: All arcmira_execute_read, arcmira_execute_write, arcmira_describe, and arcmira_feedback calls route to Arcmira servers; monitor.create and monitor.addEntities establish persistent alerting data flows.
Should I install skills/arcmira?
Oathe's verdict for skills/arcmira is SAFE with a trust score of 85/100. Recommendation: Install with caution.