List each agent workflow with its connectors, ask Claude to mark which have private data, untrusted input and an outbound channel, and fix any with all three. Keep the audit in a GitHub repository so every change is reviewed in a pull request.
Independent review. Unpaid; no partnership with the tools on this page. Based on public documentation and research. Trademarks identify their owners’ products. Disclosure.
Editorial shortlist
Compare your options.
For reviewing agent capability combinations.
Inspect untrusted input, private data and external writes together; this is not certification. Choose one for implementing and reviewing code changes; switching may need setup changes.
Editorial judgment, not a personal-use endorsement or performance score. These are alternatives for the named role, not three required stack components. Vendors did not pay for, request or approve these positions.
Inspectable file changes come before editor convenience. A lower-ranked option may fit your constraints better; compare the conditions stated on each card before choosing.
Setup and portability: Use scoped repositories, tests and human review. Switching agents does not migrate settings, credentials or model access.
Reviewed 2026-10-21. This page applies a shared comparison for “implementing and reviewing code changes” to the stated task and constraints. It is documentation-based editorial guidance, not three new hands-on tests. We weigh value for the job, distinct capabilities, replacement limits, technical effort and interoperability. Check current plan terms before adopting or switching.
Confirmed = vendor-documented, not Ion-tested. Human and AI-platform steps are proposed roles.
Human approval
You change every permission yourself.
Watch out for
An AI audit is a checklist, not a security review.
Starter prompt
For each workflow, record yes, no or unknown for access to private data, untrusted content and outbound channels, with configuration evidence. Treat unknown as unresolved. Include HTTP requests, browser navigation, files, logs and tool results as possible data paths. Recommend enforceable restrictions and tests; do not claim that a prompt or approval alone removes the risk.
Written for this workflow. Paste it into your AI platform and adjust it to your situation.
No platform-specific connection: the steps run in the tools themselves (browser, upload or built-in), with documented capabilities and human review steps. This is not an end-to-end test.
No platform-specific connection: the steps run in the tools themselves (browser, upload or built-in), with documented capabilities and human review steps. This is not an end-to-end test.
No platform-specific connection: the steps run in the tools themselves (browser, upload or built-in), with documented capabilities and human review steps. This is not an end-to-end test.
No platform-specific connection: the steps run in the tools themselves (browser, upload or built-in), with documented capabilities and human review steps. This is not an end-to-end test.
Gemini
Confirmed
No platform-specific connection: the steps run in the tools themselves (browser, upload or built-in), with documented capabilities and human review steps. This is not an end-to-end test.
No platform-specific connection: the steps run in the tools themselves (browser, upload or built-in), with documented capabilities and human review steps. This is not an end-to-end test.
What the platform statuses mean
Confirmed: documented access meets the recorded requirements. Inferred: a reviewed note describes a plausible, untested pairing. To verify: a connection or step is unresolved. MCP and API pairings need route-specific evidence; a plugin, CLI or API alone does not prove compatible access, credentials or permissions. These are documentation assessments, not IonBin runtime tests.
What we checked (2026-10-17)
Simon Willison’s lethal trifecta (June 16, 2025) is the combination of access to private data, exposure to untrusted content and the ability to communicate externally. With all three in one agent session, text an attacker controls can lead the agent to send your data out. He is deeply suspicious of guardrail products that claim to catch “95% of attacks”, because in security 95% is a failing grade, and says that for people mixing tools together the only safe option is to avoid the combination entirely. See the lethal trifecta.
Set it up
Write a table of every agent workflow: what it reads, where untrusted text can enter, and what it can send, post or open.
Ask Claude to mark the three legs for each row and suggest which leg to remove.
Remove unnecessary data access, untrusted inputs or outbound channels using enforced controls. Splitting steps helps only if the boundary prevents malicious instructions or secrets crossing with the handoff. An approval step is an additional control, not removal of a leg.
Store the audit in a restricted repository, without credentials or sensitive payloads. Configure review rules if approval must be mandatory, and test attempted disallowed actions after each permission change.
This is a useful threat-model checklist, not a complete security assessment. Missing one leg does not rule out destructive writes, credential theft or other prompt-injection harms.
Related agent skills
Two reviewed skills can support this check. Sentry’s Skill Scanner looks for skills that change agent settings or reach out to unknown sites. OpenAI’s Security Threat Model maps trust boundaries and abuse paths in a codebase. Neither is a security guarantee, and IonBin reviewed both from source only.