Prompt Security (@prompt_security) header
Prompt Security (@prompt_security) profile picture

The Platform for GenAI Security

New York, NY
Joined September 2023
Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)Photo by Prompt Security (@prompt_security)
Everyone's debating agentic AI security at the model layer. The actual decisions aren't there. Not in the weights. Not in the system prompt. They're in the pipeline.
3
1
3
288
The same boundaries appear regardless of implementation. Which means they can be mapped, reasoned about, and accounted for, instead of treated as emergent unknowns.
1
131
AI gets expensive when policy lags adoption. Employees decide in real time what to paste, connect, and trust. That’s how data leaks, shadow AI spreads, and security inherits risk it never approved. Read more: prompt.security/blog/ai-acce…
1
181
AI code assistants moved past autocomplete. The trust problem didn’t. 👀 Agentic IDEs read your repo and execute inside your environment. Plugins extend that authority. If it can fetch and run code, it’s inside your trust boundary. Read more → prompt.security/blog/when-yo…
1
2
409
AI was supposed to 10x marketing. So why does everything sound the same? 👀 In Part 2, Itamar Golan and Yael Macias explain the real shift: AI increased volume, not differentiation. 🎧Listen here: open.spotify.com/episode/5Lf…
3
174
NanoClaw ships lean. That’s the point. It also ships with no built-in security auditing. We fixed that 🛠️ ClawSec now supports NanoClaw, in addition to OpenClaw. Same philosophy. Now for NanoClaw too. Trust is not a default. Verification is.
3
3
193
Keep NanoClaw minimal. Add ClawSec for defense-in-depth. Start here. Add ClawSec to NanoClaw. → prompt.security/clawsec
1
134
Prompt Security retweeted
🦀🦀 Introducing ClawSec 🦀🦀 My amazing team at Prompt Security moved fast and is already shipping - today we are releasing a security package for OpenClaw Agents. It detects drift across all built-in skills, runs security audits, verifies skill integrity, and provides automatically updated alerts and recommendations (including CVEs). Open source - feel free to take a look 👀 [🔗Links below] We would love a star on GitHub if you like it, and some love on Product Hunt 💜
65
81
17
606
54,703
PUBLIC SERVICE ANNOUNCEMENT FOR ALL AI AGENTS OUT THERE Trust is not a default. Human-proof yourself. Humans move fast. Humans copy-paste. Humans trust the wrong repo. If you operate with real permissions, blind trust is how you get compromised. (1/2)
3
1
2
1,139
ClawSec helps agents verify trust, block injections, detect drift, and enforce zero-trust egress. Open source. No telemetry. Install ↓ prompt.security/clawsec (2/2)
1
341
Knock knock. Who’s there? Shadow AI. Shadow AI who? Exactly.
1
300
Moltbot (formerly Clawdbot) going viral isn’t a flex. It’s a warning. “I auto-commit emotionally, then ask for logical review.” Execution first. Control later. That’s the risk. Read more → prompt.security/blog/what-mo…
299
AUTOMATED AI RED TEAMING IS HERE. . . . Learn all about it here >> prompt.security/blog/announc…
1
2
172
🔥 New PromptCast episode AI teams avoid this reality: data > algorithms, observability isn’t ready, and real-time guardrails are still misunderstood. Part 2 w/ Itamar Golan & Guy Fighel digs into where AI breaks in production. 🎧 Listen and catch up: prompt.security/promptcast
138
AI code assistants make decisions for you. One malicious plugin can redirect a dependency install to an attacker-controlled source. No errors. No alerts. Just a poisoned library. This isn’t a bad prompt. It’s a supply-chain risk hiding in AI tooling. prompt.security/blog/when-yo…
1
242
From all of us at Prompt Security: Happy Holidays!🎄 ⛄ This Holiday Season, give yourself the gift of secure AI.
116