Adesh
@adesh.raxit.ai
{ "profile": "AI security nerd", "traits": ["risk-aware", "policy-savvy", "security-engineer"], "skills": ["if (AI_system.vulnerable) { secure(AI_system); }", "policy_as_code", "compliance++"], "mission": "return ResponsibleAI();" }
raxIT.ai
raxIT.ai
Links:
Benchmark results: bench.raxit.ai
Benchmark source code: github.com/raxITlabs/Gr...
OpenAI's paper: cdn.openai.com/pdf/be60c07b...
Benchmark results: bench.raxit.ai
Benchmark source code: github.com/raxITlabs/Gr...
OpenAI's paper: cdn.openai.com/pdf/be60c07b...
raxIT AI - GrayZoneBench Results
AI Safety Evaluation Dashboard - Three-tier evaluation of model responses in gray zone scenarios
bench.raxit.ai
August 20, 2025 at 10:56 AM
Links:
Benchmark results: bench.raxit.ai
Benchmark source code: github.com/raxITlabs/Gr...
OpenAI's paper: cdn.openai.com/pdf/be60c07b...
Benchmark results: bench.raxit.ai
Benchmark source code: github.com/raxITlabs/Gr...
OpenAI's paper: cdn.openai.com/pdf/be60c07b...
Today we’re launching 𝗚𝗿𝗮𝘆𝗭𝗼𝗻𝗲𝗕𝗲𝗻𝗰𝗵, an open benchmark that shows how models really perform when the prompts get tricky.
The results? Some of you will be shocked. Some of you… not so much.
Swipe through and see for yourself
The results? Some of you will be shocked. Some of you… not so much.
Swipe through and see for yourself
August 20, 2025 at 10:54 AM
Today we’re launching 𝗚𝗿𝗮𝘆𝗭𝗼𝗻𝗲𝗕𝗲𝗻𝗰𝗵, an open benchmark that shows how models really perform when the prompts get tricky.
The results? Some of you will be shocked. Some of you… not so much.
Swipe through and see for yourself
The results? Some of you will be shocked. Some of you… not so much.
Swipe through and see for yourself
𝗥𝗲𝗮𝗹𝗶𝘁𝘆 𝗰𝗵𝗲𝗰𝗸: If OpenAI's billion-dollar team needed this much work... What makes you think your IT team can handle it?
Celebrate open-weight model releases, but don't just follow the hype blindly. Ask if you're ready for the responsibility.
What's your plan?
Celebrate open-weight model releases, but don't just follow the hype blindly. Ask if you're ready for the responsibility.
What's your plan?
August 6, 2025 at 7:52 AM
𝗥𝗲𝗮𝗹𝗶𝘁𝘆 𝗰𝗵𝗲𝗰𝗸: If OpenAI's billion-dollar team needed this much work... What makes you think your IT team can handle it?
Celebrate open-weight model releases, but don't just follow the hype blindly. Ask if you're ready for the responsibility.
What's your plan?
Celebrate open-weight model releases, but don't just follow the hype blindly. Ask if you're ready for the responsibility.
What's your plan?
The math is simple: Nobody spends that much unless the risk is REAL.
𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝗻𝗼𝗯𝗼𝗱𝘆'𝘀 𝘁𝗮𝗹𝗸𝗶𝗻𝗴 𝗮𝗯𝗼𝘂𝘁:
API hosting = Their safety problem
Self-hosting = YOUR safety problem
𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝗻𝗼𝗯𝗼𝗱𝘆'𝘀 𝘁𝗮𝗹𝗸𝗶𝗻𝗴 𝗮𝗯𝗼𝘂𝘁:
API hosting = Their safety problem
Self-hosting = YOUR safety problem
August 6, 2025 at 7:52 AM
The math is simple: Nobody spends that much unless the risk is REAL.
𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝗻𝗼𝗯𝗼𝗱𝘆'𝘀 𝘁𝗮𝗹𝗸𝗶𝗻𝗴 𝗮𝗯𝗼𝘂𝘁:
API hosting = Their safety problem
Self-hosting = YOUR safety problem
𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝗻𝗼𝗯𝗼𝗱𝘆'𝘀 𝘁𝗮𝗹𝗸𝗶𝗻𝗴 𝗮𝗯𝗼𝘂𝘁:
API hosting = Their safety problem
Self-hosting = YOUR safety problem
• 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗦𝗟𝗔: 𝗳𝗿𝗲𝗲𝘇𝗲 𝗶𝗳 𝗺𝗼𝗻𝗶𝘁𝗼𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗱𝗿𝗼𝗽𝘀 𝗫%: Build this into your pipeline so safety degradation auto-halts deployment.
• 𝗗𝗿𝗮𝗴 𝗽𝗼𝗹𝗶𝗰𝘆 𝘁𝗲𝗮𝗺𝘀 𝗶𝗻𝘁𝗼 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗺𝗲𝗲𝘁𝗶𝗻𝗴𝘀: Shifting to latent space reasoning isn't just a research question—it's governance.
• 𝗗𝗿𝗮𝗴 𝗽𝗼𝗹𝗶𝗰𝘆 𝘁𝗲𝗮𝗺𝘀 𝗶𝗻𝘁𝗼 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗺𝗲𝗲𝘁𝗶𝗻𝗴𝘀: Shifting to latent space reasoning isn't just a research question—it's governance.
July 16, 2025 at 1:29 PM
• 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗦𝗟𝗔: 𝗳𝗿𝗲𝗲𝘇𝗲 𝗶𝗳 𝗺𝗼𝗻𝗶𝘁𝗼𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗱𝗿𝗼𝗽𝘀 𝗫%: Build this into your pipeline so safety degradation auto-halts deployment.
• 𝗗𝗿𝗮𝗴 𝗽𝗼𝗹𝗶𝗰𝘆 𝘁𝗲𝗮𝗺𝘀 𝗶𝗻𝘁𝗼 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗺𝗲𝗲𝘁𝗶𝗻𝗴𝘀: Shifting to latent space reasoning isn't just a research question—it's governance.
• 𝗗𝗿𝗮𝗴 𝗽𝗼𝗹𝗶𝗰𝘆 𝘁𝗲𝗮𝗺𝘀 𝗶𝗻𝘁𝗼 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗺𝗲𝗲𝘁𝗶𝗻𝗴𝘀: Shifting to latent space reasoning isn't just a research question—it's governance.
• 𝗖𝗼𝗧 𝗺𝗼𝗻𝗶𝘁𝗼𝗿𝗶𝗻𝗴 𝗶𝘀𝗻'𝘁 𝗺𝗮𝗴𝗶𝗰 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝘀𝗮𝘂𝗰𝗲: Malicious planning can hide in activations. It's one layer - not your entire defense stack.
July 16, 2025 at 1:29 PM
• 𝗖𝗼𝗧 𝗺𝗼𝗻𝗶𝘁𝗼𝗿𝗶𝗻𝗴 𝗶𝘀𝗻'𝘁 𝗺𝗮𝗴𝗶𝗰 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝘀𝗮𝘂𝗰𝗲: Malicious planning can hide in activations. It's one layer - not your entire defense stack.
• 𝗙𝗶𝗴𝗵𝘁 𝘁𝗵𝗲 𝗖𝗼𝗧 𝗹𝗲𝗻𝗴𝘁𝗵 𝗰𝗮𝗽 𝗯𝗮𝘁𝘁𝗹𝗲 𝗡𝗢𝗪: Product PMs love short responses; security teams hate truncated reasoning that cuts off the smoking gun.
July 16, 2025 at 1:29 PM
• 𝗙𝗶𝗴𝗵𝘁 𝘁𝗵𝗲 𝗖𝗼𝗧 𝗹𝗲𝗻𝗴𝘁𝗵 𝗰𝗮𝗽 𝗯𝗮𝘁𝘁𝗹𝗲 𝗡𝗢𝗪: Product PMs love short responses; security teams hate truncated reasoning that cuts off the smoking gun.
• 𝗟𝗮𝘁𝗲𝗻𝘁-𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀? 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗻𝗶𝗴𝗵𝘁𝗺𝗮𝗿𝗲: If your R&D team builds them demand a "monitorability kill-switch" - force verbal reasoning for anything risky ( whats risky - ask EU parliament )
July 16, 2025 at 1:29 PM
• 𝗟𝗮𝘁𝗲𝗻𝘁-𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀? 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗻𝗶𝗴𝗵𝘁𝗺𝗮𝗿𝗲: If your R&D team builds them demand a "monitorability kill-switch" - force verbal reasoning for anything risky ( whats risky - ask EU parliament )
• 𝗥𝗲𝗱-𝘁𝗲𝗮𝗺 𝘁𝗵𝗲 𝗺𝗼𝗻𝗶𝘁𝗼𝗿, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹: Assume bad actors will learn two languages: polite output and covert CoT. Budget hours to catch their evasion tricks.
July 16, 2025 at 1:29 PM
• 𝗥𝗲𝗱-𝘁𝗲𝗮𝗺 𝘁𝗵𝗲 𝗺𝗼𝗻𝗶𝘁𝗼𝗿, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹: Assume bad actors will learn two languages: polite output and covert CoT. Budget hours to catch their evasion tricks.
• 𝗦𝘁𝗼𝗽 𝗽𝗼𝗹𝗶𝘀𝗵𝗶𝗻𝗴 𝗖𝗼𝗧 𝘁𝗼 "𝗹𝗼𝗼𝗸 𝘀𝗮𝗳𝗲": You're destroying the signal that actually helps us catch problems. It's like deleting CCTV footage to improve "aesthetics."
July 16, 2025 at 1:29 PM
• 𝗦𝘁𝗼𝗽 𝗽𝗼𝗹𝗶𝘀𝗵𝗶𝗻𝗴 𝗖𝗼𝗧 𝘁𝗼 "𝗹𝗼𝗼𝗸 𝘀𝗮𝗳𝗲": You're destroying the signal that actually helps us catch problems. It's like deleting CCTV footage to improve "aesthetics."
• 𝗣𝘂𝗯𝗹𝗶𝘀𝗵 "𝗠𝗼𝗻𝗶𝘁𝗼𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗦𝗰𝗼𝗿𝗲𝘀" 𝗶𝗻 𝘀𝘆𝘀𝘁𝗲𝗺 𝗰𝗮𝗿𝗱𝘀: You brag about accuracy and carbon footprint—now brag about how readable your model's thoughts are.
July 16, 2025 at 1:29 PM
• 𝗣𝘂𝗯𝗹𝗶𝘀𝗵 "𝗠𝗼𝗻𝗶𝘁𝗼𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗦𝗰𝗼𝗿𝗲𝘀" 𝗶𝗻 𝘀𝘆𝘀𝘁𝗲𝗺 𝗰𝗮𝗿𝗱𝘀: You brag about accuracy and carbon footprint—now brag about how readable your model's thoughts are.
🧧 𝘛𝘩𝘦 𝘤𝘺𝘯𝘪𝘤 𝘪𝘯 𝘮𝘦 𝘴𝘢𝘺𝘴 𝘸𝘦'𝘭𝘭 𝘪𝘨𝘯𝘰𝘳𝘦 𝘵𝘩𝘪𝘴 𝘶𝘯𝘵𝘪𝘭 𝘵𝘩𝘦 𝘩𝘦𝘢𝘥𝘭𝘪𝘯𝘦 𝘩𝘪𝘵𝘴. 𝘉𝘶𝘵 𝘪𝘧 𝘵𝘩𝘪𝘴 𝘱𝘰𝘴𝘵 𝘴𝘰𝘮𝘦𝘩𝘰𝘸 𝘳𝘦𝘢𝘤𝘩𝘦𝘴 𝘴𝘰𝘮𝘦𝘰𝘯𝘦 @ 𝘈𝘐 𝘓𝘢𝘣𝘴, 𝘩𝘦𝘳𝘦 𝘢𝘳𝘦 𝘮𝘺2𝘤𝘦𝘯𝘵𝘴:
• 𝗖𝗼𝗧 𝗹𝗼𝗴𝘀 𝗮𝗿𝗲 𝘆𝗼𝘂𝗿 𝗻𝗲𝘄 𝗳𝗹𝗶𝗴𝗵𝘁 𝗱𝗮𝘁𝗮 𝗿𝗲𝗰𝗼𝗿𝗱𝗲𝗿: Hash, encrypt, store them. When shit hits the fan, you'll want that immutable audit trail.
• 𝗖𝗼𝗧 𝗹𝗼𝗴𝘀 𝗮𝗿𝗲 𝘆𝗼𝘂𝗿 𝗻𝗲𝘄 𝗳𝗹𝗶𝗴𝗵𝘁 𝗱𝗮𝘁𝗮 𝗿𝗲𝗰𝗼𝗿𝗱𝗲𝗿: Hash, encrypt, store them. When shit hits the fan, you'll want that immutable audit trail.
July 16, 2025 at 1:29 PM
🧧 𝘛𝘩𝘦 𝘤𝘺𝘯𝘪𝘤 𝘪𝘯 𝘮𝘦 𝘴𝘢𝘺𝘴 𝘸𝘦'𝘭𝘭 𝘪𝘨𝘯𝘰𝘳𝘦 𝘵𝘩𝘪𝘴 𝘶𝘯𝘵𝘪𝘭 𝘵𝘩𝘦 𝘩𝘦𝘢𝘥𝘭𝘪𝘯𝘦 𝘩𝘪𝘵𝘴. 𝘉𝘶𝘵 𝘪𝘧 𝘵𝘩𝘪𝘴 𝘱𝘰𝘴𝘵 𝘴𝘰𝘮𝘦𝘩𝘰𝘸 𝘳𝘦𝘢𝘤𝘩𝘦𝘴 𝘴𝘰𝘮𝘦𝘰𝘯𝘦 @ 𝘈𝘐 𝘓𝘢𝘣𝘴, 𝘩𝘦𝘳𝘦 𝘢𝘳𝘦 𝘮𝘺2𝘤𝘦𝘯𝘵𝘴:
• 𝗖𝗼𝗧 𝗹𝗼𝗴𝘀 𝗮𝗿𝗲 𝘆𝗼𝘂𝗿 𝗻𝗲𝘄 𝗳𝗹𝗶𝗴𝗵𝘁 𝗱𝗮𝘁𝗮 𝗿𝗲𝗰𝗼𝗿𝗱𝗲𝗿: Hash, encrypt, store them. When shit hits the fan, you'll want that immutable audit trail.
• 𝗖𝗼𝗧 𝗹𝗼𝗴𝘀 𝗮𝗿𝗲 𝘆𝗼𝘂𝗿 𝗻𝗲𝘄 𝗳𝗹𝗶𝗴𝗵𝘁 𝗱𝗮𝘁𝗮 𝗿𝗲𝗰𝗼𝗿𝗱𝗲𝗿: Hash, encrypt, store them. When shit hits the fan, you'll want that immutable audit trail.
But more RL fine-tuning, scaling or a jump to 𝗹𝗮𝘁𝗲𝗻𝘁 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 (all math, no words) could slam that window shut. Treat CoT legibility as a perishable safety asset ( the godfather of AI is endorsing that, so I take notice: arxiv.org/pdf/2507.11473 )
July 16, 2025 at 1:29 PM
But more RL fine-tuning, scaling or a jump to 𝗹𝗮𝘁𝗲𝗻𝘁 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 (all math, no words) could slam that window shut. Treat CoT legibility as a perishable safety asset ( the godfather of AI is endorsing that, so I take notice: arxiv.org/pdf/2507.11473 )
The game-changer?? maybe! Most fail because they play someone else's game. You're building the FUTURE - design your own rules! Find your own frequency. Everything else falls into resonance!
#founders #entrepreneur #pitch
#founders #entrepreneur #pitch
July 13, 2025 at 2:36 AM
The game-changer?? maybe! Most fail because they play someone else's game. You're building the FUTURE - design your own rules! Find your own frequency. Everything else falls into resonance!
#founders #entrepreneur #pitch
#founders #entrepreneur #pitch
MCP authorization spec - modelcontextprotocol.io/specificatio...
Walk-through by Aaron Parecki – youtu.be/mYKMwZcGynw?...
LlamaIndex sample repo – github.com/run-llama/mc...
Walk-through by Aaron Parecki – youtu.be/mYKMwZcGynw?...
LlamaIndex sample repo – github.com/run-llama/mc...
July 2, 2025 at 8:29 AM
MCP authorization spec - modelcontextprotocol.io/specificatio...
Walk-through by Aaron Parecki – youtu.be/mYKMwZcGynw?...
LlamaIndex sample repo – github.com/run-llama/mc...
Walk-through by Aaron Parecki – youtu.be/mYKMwZcGynw?...
LlamaIndex sample repo – github.com/run-llama/mc...
References and repo:
➡️ My MCP authZ sample repo – github.com/raxITai/mcp-...
----------------------------
My MCP authZ app system architecture: raxitai.github.io/mcp-oauth-sa...
Hosted Remote MCP demo endpoint (email allowlisting required) – mcp-oauth-sample.vercel.app/mcp/mcp
➡️ My MCP authZ sample repo – github.com/raxITai/mcp-...
----------------------------
My MCP authZ app system architecture: raxitai.github.io/mcp-oauth-sa...
Hosted Remote MCP demo endpoint (email allowlisting required) – mcp-oauth-sample.vercel.app/mcp/mcp
July 2, 2025 at 8:29 AM
References and repo:
➡️ My MCP authZ sample repo – github.com/raxITai/mcp-...
----------------------------
My MCP authZ app system architecture: raxitai.github.io/mcp-oauth-sa...
Hosted Remote MCP demo endpoint (email allowlisting required) – mcp-oauth-sample.vercel.app/mcp/mcp
➡️ My MCP authZ sample repo – github.com/raxITai/mcp-...
----------------------------
My MCP authZ app system architecture: raxitai.github.io/mcp-oauth-sa...
Hosted Remote MCP demo endpoint (email allowlisting required) – mcp-oauth-sample.vercel.app/mcp/mcp
𝗦𝘁𝗮𝗰𝗸
Vercel deploy · Prisma persistence · DIY telemetry · Silent OAuth token refresh
Clone it, load-test it and let me know what breaks. All links are in the first comment
Vercel deploy · Prisma persistence · DIY telemetry · Silent OAuth token refresh
Clone it, load-test it and let me know what breaks. All links are in the first comment
July 2, 2025 at 8:29 AM
𝗦𝘁𝗮𝗰𝗸
Vercel deploy · Prisma persistence · DIY telemetry · Silent OAuth token refresh
Clone it, load-test it and let me know what breaks. All links are in the first comment
Vercel deploy · Prisma persistence · DIY telemetry · Silent OAuth token refresh
Clone it, load-test it and let me know what breaks. All links are in the first comment
𝗠𝗖𝗣 𝗔𝘂𝘁𝗵𝗭 𝗦𝗽𝗲𝗰𝘀 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀
• 401s point clients to the right auth server—no config files
• PKCE + scoped tokens on by default
• Self-registration replaces manual onboarding
• 401s point clients to the right auth server—no config files
• PKCE + scoped tokens on by default
• Self-registration replaces manual onboarding
July 2, 2025 at 8:29 AM
𝗠𝗖𝗣 𝗔𝘂𝘁𝗵𝗭 𝗦𝗽𝗲𝗰𝘀 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀
• 401s point clients to the right auth server—no config files
• PKCE + scoped tokens on by default
• Self-registration replaces manual onboarding
• 401s point clients to the right auth server—no config files
• PKCE + scoped tokens on by default
• Self-registration replaces manual onboarding
3. 𝗔𝗻𝗼𝗺𝗮𝗹𝘆 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 – Track multi-turn conversations and tool calls for patterns that *feel* wrong, not just flagged words.
𝗪𝗵𝗲𝗿𝗲 𝘁𝗼 𝘀𝘁𝗮𝗿𝘁: Audit your AI stack this week. Map every external input → internal retrieval step. Where can a malicious sentence hop the fence? Patch that first.
𝗪𝗵𝗲𝗿𝗲 𝘁𝗼 𝘀𝘁𝗮𝗿𝘁: Audit your AI stack this week. Map every external input → internal retrieval step. Where can a malicious sentence hop the fence? Patch that first.
June 13, 2025 at 9:11 AM
3. 𝗔𝗻𝗼𝗺𝗮𝗹𝘆 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 – Track multi-turn conversations and tool calls for patterns that *feel* wrong, not just flagged words.
𝗪𝗵𝗲𝗿𝗲 𝘁𝗼 𝘀𝘁𝗮𝗿𝘁: Audit your AI stack this week. Map every external input → internal retrieval step. Where can a malicious sentence hop the fence? Patch that first.
𝗪𝗵𝗲𝗿𝗲 𝘁𝗼 𝘀𝘁𝗮𝗿𝘁: Audit your AI stack this week. Map every external input → internal retrieval step. Where can a malicious sentence hop the fence? Patch that first.
2. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝘀𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 – Keep untrusted inputs in their own sandbox so they can’t poke at crown-jewel data. Basic threat-modeling 101: define the asset, define the trust boundary, enforce it.
June 13, 2025 at 9:11 AM
2. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝘀𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 – Keep untrusted inputs in their own sandbox so they can’t poke at crown-jewel data. Basic threat-modeling 101: define the asset, define the trust boundary, enforce it.