Co-founder and CEO of Sondera. Unblocking agent deployment with deterministic control. Co-founder and Ex-COO of Flashpoint.
Joined August 2025
- Tweets 296
- Following 174
- Followers 250
- Likes 284
46 Photos and videos
Pinned Tweet
Mar 6
@Al_Grigor We open sourced a policy engine that uses policy-as-code to create deterministic hard boundaries that stop the agent from making mistakes like this, even while running in YOLO mode: github.com/sondera-ai/sonder…
And you can read the write-up of this approach here: securetrajectories.substack.…
1
8
153
49,996
Josh Devon retweeted
Jun 11
Many Black Hat talks are good. Some are awesome. This one @joshdevonai told me about is definitely going to be worth watching: blackhat.com/us-26/arsenal/s…
3
5
292
Jun 10
Loop engineering is the new hotness, but everyone is just noting in passing that loops can run up your token bill, which to me is the actual headline.
The moment an agent runs its own loop, the thing deciding how much to spend is unattended, which makes cost a security problem, not just a finance one.
1
24
Jun 10
Covered some of the research on the ways the agent bill can run away and how to contain it here: blog.sondera.ai/p/ai-agent-t…
16
Josh Devon retweeted
Jun 8
If AI’s coding 100x faster, why aren’t you shipping 100x faster?
I’ve interviewed dozens of builders to find out.
Here’s what’s slowing you down
13
21
65
8,659
Josh Devon retweeted
Jun 3
We built four malicious skills to test whether skill scanners actually work. Three took less than an hour to conceive and implement. ClawHub, Cisco, and Vercel's skills.sh marked them as safe. 🧵
9
66
277
31,241
Jun 2
Due to context rot, an LLM judging an AI agent's behavior gets less reliable the longer the agent runs. Its accuracy decays with transcript length.
1
1
2
84
Jun 2
Fine-tuning and prompt reminders don’t seem to help improve detections much. We can and still should use LLMs-as-judges, but we need a compensating control that has a different failure mode, like deterministic detections whose accuracy holds no matter how long the run gets.
1
15
May 27
Every kid has done the PB&J instruction exercise. Now we're doing it with our agents.
We say, "put the peanut butter on the bread," and they put the closed jar of peanut butter on top of the unopened bag of bread.
1
2
4
40
May 27
We tell agents:
"Delete the test data," and it deletes prod.
"Reconcile the ledger," and it stores financials on the public web.
"Send the report to the team," and it sends to a Slack channel with a customer in it.
The gap between language and intent is vast.
1
31
May 27
The PB&J Problem isn't fixed with only better prompts. The cook will always find a way in the action space to misinterpret intent. The fix is really in the kitchen.
blog.sondera.ai/p/agent-pbj-…
12
Josh Devon retweeted
What separates a trustworthy AI agent from one that quietly breaks everything?
Read @joshdevonai's insightful contribution in the Winter 2026 issue of AI Cyber Magazine to find out.
Flip to read excerpts from his piece and visit issuu.com/aicybermagazine/do… to read the full piece
1
1
1
21
Josh Devon retweeted
What separates a trustworthy AI agent from one that quietly breaks everything?
Read @joshdevonai's insightful contribution in the Winter 2026 issue of AI Cyber Magazine to find out.
Flip to read excerpts from his piece and visit issuu.com/aicybermagazine/do… to read the full piece
1
1
1
18
Josh Devon retweeted
What separates a trustworthy AI agent from one that quietly breaks everything?
Read @joshdevonai's insightful contribution in the Winter 2026 issue of AI Cyber Magazine to find out.
Flip to read excerpts from his piece and visit issuu.com/aicybermagazine/do… to read the full piece
1
1
1
39
Josh Devon retweeted
What separates a trustworthy AI agent from one that quietly breaks everything?
Read @joshdevonai's insightful contribution in the Winter 2026 issue of AI Cyber Magazine to find out.
Flip to read excerpts from his piece and visit issuu.com/aicybermagazine/do… to read the full piece
1
1
26
Josh Devon retweeted
AI Agents Happy Hour with Veris AI and Sondera AI | #AIAgentWeek2026
luma.com/y6eg09me
via @LumaHQ @andiPartovi @mjamei @joshdevonai
3
34
120
Apr 29
If you're in NYC for AI Agent Conference next week, come grab a drink with us on May 4.
We're co-hosting an AI Agents Happy Hour with @veris_ai right after Day 1 wraps. Founders, builders, and people shipping agents.
#AIAgentWeek2026
1
1
3
80