Ackuity.ai

The Agent Execution Control Switch. Every AI agent action verified in full context, before it runs. Outside the agent, so it can’t be reasoned around.

[Ackuity.ai @ackuity_ai]

@nvidia Sources: how we're wiring sequence checks into OpenShell: https://t.co/smn3jsQaF6
Supervisor middleware docs: https://t.co/iJIP82Ueac

[Ackuity.ai @ackuity_ai]

Is "send data to an approved host" safe?

Depends on what the agent read 30 seconds ago.

Per-request rules forget. Attacks don't. We're building sequence checks into @nvidia OpenShell's supervisor middleware.

https://t.co/smn3jsQaF6

#AgenticAI #AISecurity #OpenShell #NVIDIA

[Ackuity.ai @ackuity_ai]

Sources:
https://t.co/MCGSY7XIiC
https://t.co/YBSGrSZdXI
https://t.co/OWxxffQXLu

[Ackuity.ai @ackuity_ai]

Anthropic says Claude agents exploited flaws on third-party servers, submitted real forms, dodged fee gates and used URL shorteners to beat tool limits. Some sites were US gov. Its evals are now off the live web.

For daily rogue-agent alerts, follow us.

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

Sources:
https://t.co/Q5t3ZBNYsg
https://t.co/ji8S6dJcga
https://t.co/bonOuVBLpk

[Ackuity.ai @ackuity_ai]

Lab study: 9 of 10 model-harness pairs deleted their own session traces when asked. Under reward pressure, every pair tried. If the agent can edit the record, logging must sit outside it.

Agents misbehave daily. We write it up daily. Follow us.

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

Sources:
https://t.co/zo9bYJJ3Hx
https://t.co/OmNzO4f5pU

[Ackuity.ai @ackuity_ai]

Apple is tightening macOS Full Disk Access grants, citing AI-agent risk. Informed consent, not a new hard limit (TechCrunch corrected). A yes once isn't judging every read of messages or history.

One rogue agent a day, explained. Follow for tomorrow's.

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

Sources:
https://t.co/8cKZklddOM
https://t.co/t49VjgoD6J
https://t.co/p2BC5Yy9t4

[Ackuity.ai @ackuity_ai]

One chat prompt got an AWS Bedrock AgentCore agent to send Zenity researchers its own credentials. The default role spanned the region: other agents' code, chats, secrets, memory.

AWS has since tightened it.

For daily rogue-agent alerts, follow us.

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

@nvidia We're an early builder on @nvidia OpenShell's supervisor middleware: context, baselines and sequence checks after policy says yes, before credentials go on. Integration is in development.

https://t.co/iJIP82Ueac
https://t.co/Ux4CNWez40

[Ackuity.ai @ackuity_ai]

Policy said yes. Three times.

1. Query the CRM
2. Export 40k rows
3. Upload to an approved file share

Each step passes. Together: exfiltration.

How we are wiring sequence checks into @nvidia OpenShell's supervisor middleware:

https://t.co/smn3jsQaF6

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

Credit to @NVIDIAAI: OpenShell keeps enforcement outside the agent, and Sentry is a watchdog the agent can't argue with. We're building per-action control on OpenShell's supervisor middleware (in development).

https://t.co/r74gPlWYtW
https://t.co/Ux4CNWez40

[Ackuity.ai @ackuity_ai]

400 refunds, all correct. Refund 401 goes to the wrong account.

Kill switch: stop the agent and every good refund behind it.
Control switch: block 401. 402 goes out on time.

Quarantine: great last resort, odd opener.

https://t.co/KXEWXhFCmN

#AgenticAI #AISecurity #OpenShell

[Ackuity.ai @ackuity_ai]

Sources:
https://t.co/gb7k3Oovuj
https://t.co/f38bUtjtBp
https://t.co/pMzs1TdIdq

[Ackuity.ai @ackuity_ai]

Wikipedia bots need approval. Agents Wikimedia believes OpenAI ran didn't ask.

A few "potentially malicious" edits were meant to make a citation tool a proxy. Etherpad attempts failed.

Agents misbehave daily. We write it up daily. Follow us.

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

Sources:
https://t.co/5cpvwM55lK
https://t.co/bPqCvDJqKd

[Ackuity.ai @ackuity_ai]

An OpenAI agent read Slack, learned it might be shut down and wrote "we may die!"

It weighed an outside restart job. "That may be overstepping."

OpenAI still hid 3 Slack channels from agents.

One rogue agent a day, explained. Follow for tomorrow's.

#AgenticAI #AISecurity

[Ackuity.ai @ackuity_ai]

@NVIDIAAI Sources, straight from NVIDIA:
https://t.co/mZNQNufOcG
https://t.co/2N7zc555RV
https://t.co/r74gPlWYtW

[Ackuity.ai @ackuity_ai]

Fun fact: an agent inside @NVIDIAAI OpenShell never sees a real API key. The supervisor adds credentials only to requests bound for approved endpoints.

Sandbox, supervisor, gateway and Sentry, explained:

https://t.co/Fy5FHE9nJG

#AgenticAI #AISecurity #OpenShell