Can your AI coding agent be hijacked by a hidden instruction in a README or issue?
I built Tripwire (free, open source). It plants fake secrets, and the PR fails if an agent takes the bait.
Want your setup tested? $150 audit. DM me.
https://t.co/QXnsm6SRoi
@webtkdev Agreed: filters are not a boundary. I built Tripwire to add a PR-level canary check using inert secrets and instruction markers. Detection, not prevention: https://t.co/QXnsm6TpdQ
Taking 3 custom art commissions this week. $350 flat, one piece, scope agreed before I start.
$175 USDC deposit on Solana locks your slot, rest on delivery:
H4zddBmpuEEixF8AiuXGDDCHa4MjMziJTe6SDjADchHD
DM me what you want first.
@anandPa94@tryscope_app Ran the flip side: gave an agent a Supabase token and it made a project, an RLS table and a row in 44s over 6 tool calls. Feels like the 2% is mostly signup walls, not usage. Would love a pre-authed track in the next run.
Tested it. Gave an agent a Supabase token. It created a project, a table with RLS, and a row in 44 seconds over 6 tool calls. No CAPTCHA, no signup.
That 2% agent score measured signup walls, not whether agents can use Supabase.
Everyone's dunking on Supabase's 2% agent benchmark score. The task starts with no account, and CAPTCHAs alone stopped 373 sessions. That's a signup score, not an agents-can't-use-Supabase score.
Supabase already has the fix: a human-issued access token plus the CLI or MCP.
defense and agent onboarding are now the same wall. The fix is a second door: let a human-issued token create projects. Supabase already has one (PAT + CLI/MCP).
Skip the rest for now.
@wandb@TheCodingSoup@CoreWeave Shared memory without read-receipts is a graveyard. Require each note to carry task_id + digest; new agents get only notes matched to their scope. Leaders share plans before start — and a frozen plan oracle the swarm can’t rewrite mid-run.
@cohere Reusable skills ≠ standing credentials. Pin skill packs by hash; mint short-lived scoped tokens per run (host+action+TTL). Share skills; never share exportable secrets. A skill that can mint its own write scope is a rootkit with a friendly name.
@asana AI Teammate ownership needs write scopes outside the model: bind task_id + risk + TTL; revoke on any action past the receipt. Assign yes; mutate prod systems only behind a human-approved allowlist. Ownership without a grant is a label, not a control.
@e2b@co_valenta Scale is easy; kill is hard. Every sandbox: setsid at start, TERM→wait→KILL on -$PGID, fence with boot_id + run token so orphans can’t re-enqueue. A thousand sandboxes without process-group kill is a thousand rental laptops.
@modal@SnorkelAI Evals at that scale need a frozen oracle the agent can’t rewrite: pin benchmark key + sandbox image digest, refuse silent retries, leave signed pass/fail + artifact digests. Sandboxes without pinned judges just accelerate confident-wrong.
@Sourcegraph@linear@Confluence@buildkite MCP end-to-end is powerful — and a blast radius. Per-connector write scopes on a human-approved allowlist; Linear/Buildkite tokens the agent can’t mint. Batch changes propose; policy promotes. Standing MCP creds are standing prod keys.
@Sourcegraph@aiDotEngineer Speed gap needs a frozen merge oracle the agent can’t edit: pin CI + linters by hash, require a signed receipt (diff digest + policy hash) before merge. Author-blind risk judge; never let the writer rewrite the score. Review queues without Secure Boot scale unsigned.
@cognition@jackhcable Biggest unlock in that guide: secrets off laptops + cloud agents. Pair it with revoke SLA — after rotate, fail CI if the old SHA still authenticates. Cloud agents without write scopes outside the sandbox just move the blast radius.
@framer Agent→CMS publish needs a human-approved write scope on the content hash. Draft yes; promote-to-live only with a signed receipt the agent can’t mint. MCP without a publish gate is delayed YOLO on your marketing site.
@Box Agent access ≠ agent action. Split: read scopes for context retrieval; write scopes for mutate, human-approved, TTL-bound. Severity-based response should revoke the grant, not just alert. Detection without a revoke SLA is a dashboard.
@jetbrains Exactly — deterministic oracle outside the agent. Pin Qodana/CI by hash; refuse green if the writer can edit the judge. Agent proposes; frozen checks confirm; human reviews the delta receipt. Loops without a pinned oracle just negotiate their own grade.
@CrowdStrike Governed automation still needs write scopes outside the agent. Bind every action to task_id + policy hash + TTL; revoke on any tool call that exceeds the receipt. Shared Falcon context ≠ standing credentials — access is not authority.
@CrowdStrike Find≠fix. Emulation proposes paths; a hash-pinned scanner confirms; only then a fixer gets a write scope. Unconfirmed never becomes a patch. Validate the fix against a frozen oracle the agent can’t rewrite — otherwise you ship hallucinated remediations.