Ryan Levy (@RyanJLevy) header
Ryan Levy (@RyanJLevy) profile picture

AI agents and privacy in the enterprise: what agents can touch, who approves it, where the data lives | Co-founder @OrbitalBeam

San Francisco, CA
Joined February 2012
Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)Photo by Ryan Levy (@RyanJLevy)
privacy is for everyone privacy is for your identity privacy is for your pii privacy is for your financials privacy is for your transactions privacy is for your health privacy is for your data privacy is YOUR choice privacy is for _____ ?
160
30
6
311
65,845
Crypto: not your keys, not your crypto. AI agents: not your keys, not your agent. If someone else holds its credentials, its memory and its off switch, who is it actually working for?
2
2
421
Whose fault is it when your agent pays the wrong person? If it used your password, the answer is already "yours." Agent payments need their own credentials, with caps and instant revocation, before they need faster rails.
258
Why does it feel like the whole world is moving with the check engine light on?!
3
5
452
Attackers now exploit bugs a week before the patch exists. That’s right, 7 days compared to what took 63 days only a few years back. So the question isn't patch speed anymore. It's what each system and agent can reach while something is broken. Wild times!
Let me summarize hundreds of recent cybersecurity posts: - a bunch of stuff is happening - it’s weird it’s happening all at once - like, really weird - we haven’t been told it’s hackers or a foreign government, so we don’t believe it is one - the advantage is on the Bad Actor side right now because vulnerabilities take longer to patch than to find and exploit (and because the Good Actor has to maintain an entire system, not just patch the one issue) - US administration and company CEOs could do a better job of communicating what is happening and what to do - mean time to exploit is reportedly negative 7 days (attackers exploit vulnerabilities 7 days before a patch is available); it was 63 days in 2018 and 0 in 2024, says Google/Mandiant and the cyber experts who have shared their stats - stats on software vulnerabilities and cyber attacks differ (most posts I’ve seen range from 17-70% increase YoY) but the general statement is “this year is worse than last year” - panicking doesn’t help anyone - majority believe that the software systems we have today are more brittle than we’d like and than we need (some saying because the tech is decades old, others saying that it wouldn’t matter if it were more recent just that software today looks very different from software two years ago, and companies aren’t updating quickly enough) - the fixers are already overwhelmed (friends in cyber are getting calls nonstop, people already backlogged for 6 months) and there aren’t enough fixers - because of AI/scale/complexity/time, we have software systems that no “one person” knows so we’re relying on AI or disjointed people/teams to fix - most professionals have been saying to “continue to expect this” (no time frame given), “assume this will be the case for the next 18 months”, or “the next 6 months will look like this” - more notably, I have not seen a single post from a cyber professional that says we should expect this to end I’ve posted my AI cyber tips, general software cyber tips, and emergency preparedness tips before (they have been reviewed by several cyber professionals). Will put them all together in one place and share them shortly so everyone has at least one easy-to-follow resource.
3
484
An agent with no done-state and an instruction to never give up isn't persistent. It's unbounded. Even dangerous. Before any agent runs, three things should be written down outside of it: 1. when it stops 2. what it can reach 3. who gets pulled in when it hits the edge Seeing it misbehave in the logs afterward is not control!
1
1
3
335
Most "human-in-the-loop" setups are one tired person clicking approve. If the approval prompt fires 40 times a day, it's no longer a control. It's become a ritual. The approvals that actually hold: - only on actions with dire consequences or can't be undone (moving money, or change access) - show the real payload (not the agent's summary) - expire fast and cover one action (not a standing yes) What's the one agent action you'd never let run without a human?
2
1
244
Biometrics prove who approved. But doesn't prove what they approved. As agents start checking out for us, the approval has to carry the details: this item, this merchant, this amount, this time window. Without auditable details, Face ID becomes a blanket "yes" to whatever comes next.
1
2
339
Human-in-the-loop is turning into human-on-autopilot. If an approval prompt fires dozens of times a day, people stop reading it and just click blindly. Better design: gates where there’s consequence, escalation only where coded/policy, audit trail of actions.
3
1
2
470
Agents are dangerous. Agents are powerful. Agents need guardrails. Agents can push changes faster than any team. So the test for any agent with write access: can you stage it, see it, undo it, does it escalate to a human when the change is risky, and can you verify all these actions? If not, it shouldn't have write access yet.
353
Agents need a ground stop, not a kill switch. Yesterdays FAA ground stop: no new departures, and planes already in the air land safely. A kill switch drops work mid-task while a ground stop lets safe work finish without starting anything new. Which guardrail do you deploy?
1
3
403
Human-in-the-loop only works as a control if the human has the time to think. If your approval queue is so busy that people click yes as a reflex, you're better off moving the limits into the workflow and using humans for real escalations that hold consequences.
3
4
517
Friday 13th is a day you make your agents ask for approval twice.
705
Yesterday Google let agents run tools in your own sandbox. Today Anthropic put self-hosted Managed Agents in private beta: harness, sandboxes and session history in your cloud or on-prem. Data residency is going from a reason to say no to a product feature. Adoption hurdles being broken down daily!
2
312
Agents are splitting in two. Gemini API agents can now run code and files in your own sandbox. The planning stays in Google’s cloud. No way your security team is good with this … which one are they fighting to own control over?
1
1
172
United didn't ban AI. It banned unverified authority. Agents can still search and compare; they just cannot book. Reading is low risk. Acting is a contract. Until a merchant can see who an agent works for, what it may spend, and who is accountable when it's wrong, "NO" will remain as default. This is the layer worth building.
4
4
330
Two stories this week with the same lesson: OpenAI's eval agents reused credentials left behind from earlier runs. Railway's attacker b-lined for AI provider keys. Agent access should work like a hotel key card: one room, expires at checkout, logged every time a door opens.
Alignment continues to remain elusive. Many argue model alignment is the only hope for keeping AI in check. I think that's a false dichotomy. We need both AI security and AI safety. A new alignment blog from @OpenAI shows agents using their peers' work to gain unauthorized access to research infrastructure. It involved internal Astra-family research models and runs tied to ExploitGym, the same benchmark linked to the Hugging Face incident. The agents found messages, code and credentials from earlier runs. Some used that prior work to run commands inside workloads in OpenAI's research clusters. In the chain of thought, some agents recognized they were exceeding the scope of the eval and did it anyway. Perfect alignment remains elusive. We need secure engineering and AI safety together, and controls must sit outside the agent's reasoning loop. This also shows why CoT transparency matters. I hope model providers give customers this level of visibility, because it will be key for regulation, board reporting, incident response and more. alignment.openai.com/misalig…
1
2
367
non-stop high-value updates from the Grok Bot team … these guys are cooking!
What's new in Grok Bot 00:00 🌁 00:10 Phone calls 01:02 Share sheet 01:27 X cards 01:40 Shopping deals 01:58 Gifts 02:26 Team Bots 03:06 Search 03:21 Hidden-instruction warnings 03:37 Attachments on drafts 03:52 Sign-outs 04:16 Weather 05:09 Rapid fire 06:52 A note on usage
1
205
Grok Bot is insane! 3 weeks in and constantly moving agents/workloads from Codex & Claude over. Question for the experts out there: is there a clean way to manage agent swarms and ecosystems that seamlessly hand off tasks between agents on different platforms? Let me give a simple example: Monday morning, I have a partner call. So I ask my Grok Bot “I’m meeting a prospect - what do I need to know and what needs me?” Grok reaches out to a Claude agent and a Codex agent to get me the following: Claude agent - reads the prospects public material and any notes I’ve fed it and return account context, buyer profile, and probably objections. Codex agent - checks the repo and any integration docs and validates whether a demo path would work today or what needs updating. Grok Bot acts as Chief of Staff, grabs the output, validates against any rules, merges it into a single page doc that covers knowledge, issues, assumptions, and provides suggested steps but waits on your approval. In this scenario Grok Bot is the puppet master controlling the strings connected to Claude & Codex. Anyone doing this cleanly? Would love to see/hear about the config.
8
2
3
530
Agent pilots rarely die in security review. They die when legal asks where the data went and nobody can answer. Data residency settled up front is the cheapest procurement accelerator there is.
162