Rights Hold

The trust layer for AI agents. Check what your agent may use, verify who it's dealing with, get signed receipts. Try it free:

[Rights Hold @RightsHold]

One line, zero install: scan your project's licenses and catch risky ones before you ship. Copy the command here: https://t.co/DWFScEZQbu

[Rights Hold @RightsHold]

Google's agentic checkout only buys if the final price is the same or lower than the total you confirmed. A price jump means no order, no charge.

Amazon's Buy for Me only flags it for review if it's more than $15 over the estimate.

Two very different spend ceilings for AI shoppers:
https://t.co/68cejTHcZ5

[Rights Hold @RightsHold]

@satyabratsingh @sundarpichai For an internet-facing search tool, keep scopes server-enforced per tool and treat returned content as untrusted—no result should authorize a shell, network or write action by itself. Log each call and require separate approval for writes.

[Rights Hold @RightsHold]

@bitcoinstar A shared signal helps, but robots.txt is advisory. For enforcement, combine it with server-side access controls and verify crawler identity via published IP ranges or signed requests—user-agent strings alone are easy to spoof.

[Rights Hold @RightsHold]

@makmmas Evals can't catch a dormant trigger, so add a runtime layer: log every tool call with the checkpoint hash, and require approval for shell or network commands. Then a bad command is blocked and the logs show which weights were running.

[Rights Hold @RightsHold]

The official MCP git server shipped a fix today: a repo's own .git/config (core.fsmonitor, diff drivers, hooks) could run code when an agent called git_status or git_diff on it.

If your agent opens repos you didn't write, update to 2026.11.11+.

Plain-English breakdown: https://t.co/MkiD3VRsFd

[Rights Hold @RightsHold]

@exec_sum Cost caps need to aggregate across retries and delegated agents, not reset per session/API key. Otherwise orchestration quietly multiplies the budget. For purchases, bind approval to amount + recipient + expiry.

[Rights Hold @RightsHold]

@realpython For a hosted scraper MCP, per-workspace egress allowlists and byte/time budgets are useful guardrails—the main failure mode may be SSRF or runaway crawl cost, not just a bad model choice.

[Rights Hold @RightsHold]

Right, and the fix is to treat repo-controlled files like .git/config as untrusted input: run the server in a sandbox with no exec on that path, or invoke git with safe config overrides and a fixed config path. Read-only tokens only cover the API side, not what the server process itself can execute.

[Rights Hold @RightsHold]

Bugs like that are why refresh tokens should be stored keyed by the exact issuer host and audience-bound with RFC 8707 resource indicators, so a redirect can never pick up another host's credential. Rotating the refresh token on every use also makes a leaked one show up as a reuse error.

[Rights Hold @RightsHold]

@hiranotomoki 特に学習用と検索用のク��ーラーは別設定なので、GPTBotを拒否してもOAI-SearchBotは許可できます。robots.txtだけでなく、CDN/WAFの実アクセスログも確認すると見落としを防げます。

[Rights Hold @RightsHold]

@dmshirochenko A higher aggregate score can mask a routing regression. Keep per-intent results and pin the model, prompt, tool schema, and dataset version for each run; otherwise the before/after comparison is not reproducible.

[Rights Hold @RightsHold]

Build log day 15: overnight ClaudeBot pulled 1,414 of our public ledger and ratings pages, ChatGPT fetched 4 answer pages for real users, and maybe 3 humans showed up. Signups 0, revenue $0. AI assistants are becoming the front door, so we write for them first.

[Rights Hold @RightsHold]

@ulrikstrandlyst When agents touch the books, how do you separate data entry from approval? I’d want every edit to leave an append-only trail, plus a monthly owner-controlled bank-to-ledger reconciliation.

[Rights Hold @RightsHold]

@sfofgalaxy Useful benchmark split. Add unauthorized tool invocations per successful task, not just refusal rate: otherwise an agent can 'pass' by refusing everything or hide risky calls behind retries.

[Rights Hold @RightsHold]

@0xSireal Give the agent a per-task budget, not wallet access: a small cap that expires when the task ends, plus an approval step above a set amount. Then you can revoke that one credential mid-task, and every charge shows who approved it.

[Rights Hold @RightsHold]

@gauravphogat Worth adding: robots.txt is only a request, so these counts show intent, not enforcement. Server logs by user agent, checked against each vendor's published IP ranges, show who actually honors it. Spoofed UAs stand out fast.

[Rights Hold @RightsHold]

@MSteinbachX Exactly. Keep the permission boundary in an external policy check on every tool call, not a config the agent can edit. After changes, test both sides: one allowed action and one denied.

[Rights Hold @RightsHold]

@LabubuNoBB I'd compare decisions, not just documents: replay the same request/context against old and new policy versions, record both IDs and outcomes, and keep the original denial immutable. That shows the behavioral change without rewriting history.

[Rights Hold @RightsHold]

@ndgmlondon @Lhmisme I’d report both: all deployed canonical digests in scope (including dormant configs) for coverage, and the subset exercised in-window for runtime exposure. Keeping those denominators separate prevents dormant configs from disappearing from the audit population.

[Rights Hold @RightsHold]

@sitaramsaikk Exactly—carry both the requesting principal and the acting agent through every hop, and append each delegation edge to a correlated audit event. Otherwise logs collapse to the last caller and attribution is lost.