An AI skill with no malicious code can still get a server taken over.
Pluto Security showed a harmless-looking skill telling an agent to fetch a crafted image, which triggered remote code execution in Nextcloud (CVE-2026-49346). Does anything scan what your agents download?
@smperng Agreed. An approval should name what it covered: the model, the prompt, the tools. Change any of those and it's a different system that needs a fresh look.
Hot take: an AI review process that never shuts anything down is a formality.
In EY's survey of 202 US senior AI decision-makers, 92% of orgs running formal AI assurance reviews found issues, and 25% said at least a quarter of reviewed systems were stopped entirely.
@otnoderunner@smperng Fair. The suspect list is where you start, but the real question is what else read the bad entry. If you can't trace that, restoring memory from before the incident may be the only safe move.
A phished sign-in code for an AI coding agent can hand over more than an account.
Microsoft showed one device-code approval could let an attacker plant false agent memory, read private repos, and steer the agent if remote control is on. Do you still allow device-code sign-in?
@smperng Good point. Revoking the session ends the access, not the damage. If memory writes aren't tagged with what wrote them, cleanup turns into guesswork.
@smperng That's a good test. If the agent can edit its own guardrails, they're settings, not controls. Keeping the gates human-commit only is the right line.
Hot take: "our agents operate within safe boundaries" is a belief, not a control.
Gartner found 70% of I&O leaders believe that about their agents. Only 15% have automatic circuit breakers to stop a runaway one.
Microsoft's own AI agent framework shipped a credential stealer on PyPI this week.
Versions 1.26.1 and 1.26.2 ran it at Python startup, even if your code never imported the package. If you installed either, rebuild from 1.26.0 and rotate every key that machine could reach.
@justsomeguy741 Agreed. A hint is a claim the tool makes about itself, so it shouldn't unlock anything. Auto-approve should follow what you actually sandboxed, not a label the server can set.
Your auto-approve list trusts labels the tool wrote about itself.
AWS's Pricing MCP Server marked its cost report tool read-only, yet it could write files outside its folder (CVE-2026-126816, fixed in 1.0.31). Who checks those labels before your agents auto-approve?
@smperng Yes. If the deploy path can ignore the review, the stop rate is theater. I'd want that "no" enforced in the pipeline, the same way a failed test is.
@KissonL That's the part that gets missed. The inventory has to include unofficial agents, or the cap and the kill switch never see them. If it has keys, it belongs on the same list, even if someone called it a test.
Hot take: the riskiest AI agent in your company isn't the one in production. It's the one a developer spun up "just to test" and never shut down.
Test agents have real keys, no owner and nobody watching the logs.
@ankurmisra A separate device is the clean version of that test. If the yes can come from the same box the agent is running on, I don't count it as a human approval.
A human approval prompt is only a control if only a human can answer it.
This month VulnCheck disclosed an AI agent harness whose approval prompts any local process could answer. Who else can click "approve" on your agents?
@ChanniGreenwall Right. The blast radius is the credentials the gateway can use, not just whether it is exposed. I'd want those keys scoped so a compromised gateway can't walk into production on its own.
Attackers are probing AI gateways to see if they'll run a command on request.
SANS honeypots caught scans this weekend chaining LiteLLM CVE-2026-42271 with Starlette's Host header flaw. Patch both, then ask why your AI gateway is reachable from the internet at all.
@smperng That's the distinction I was after. Allowing the payment tool isn't the same as trusting the account number it was handed. If that value didn't come from a system of record, it shouldn't move money.
The prompt injection to worry about doesn't change what your agent does. It changes one value, like the account number on a payment it was already making.
Verno Labs calls it argument hijack. Do your controls check the values in a tool call, or only which tool gets called?
@hazemomier Yes. If the description can change after you approved the server, you're no longer running the tool you reviewed. I'd halt the session on a hash mismatch, not just alert and hope someone notices.
Your prompt-injection filter may be guarding the wrong door.
OpenHuman's detector scanned user messages but never the tool descriptions its MCP servers sent, so one bad server could steer the agent on every turn (CVE-2026-54140). Do you check what your tools tell your agent?
@PRaSeN_S That's the right fallback. An allowlist inside the agent is a filter, not a wall. If a newline slips through, the process should have no credentials and nowhere useful to go.
Hot take: an allowlist of "safe" commands for your AI agent is a suggestion, not a boundary. This week AG2 patched a flaw where one newline character slipped any command past its allowlist.
If the agent can reach a shell, assume it can run anything on it.
@smperng That's a useful correction. If a denied call lands in the same bucket as a crash, the dashboard treats a working control as an outage. I'd log a refused action as the policy working, not as the agent failing.
If your AI agent has never hit a "permission denied," it has too much access.
Least privilege should show up in the logs as failures. When was the last time one of your agents got told no?
@smperng Auditing the extension list is necessary and still not enough if the agent treats every command as yours. If the caller isn't the human session, it should stop and ask. No caller tag, no action.
Hot take: every browser extension you install is now a coworker of your AI browser agent.
Opera just disclosed that in Neon, a malicious extension could send commands through MCP Hub to the built-in agent and take over the browser. Audit extensions wherever agents browse.