Kelvin

Cyber security consulting... Sharing my insights on security, AI, & online business

[Kelvin @IamRealKelvin]

@DrMShujat I got to disagree… Dr Klenner did clinical research on it. Dr Levy also. Snake venoms (and most toxins) act by stealing electrons from cells, creating oxidative damage and disrupting normal redox balance. Vitamin C donates electrons as a powerful reducing agent, restore redox.

[Kelvin @IamRealKelvin]

@DrMShujat Vitamin C IV megadose. That’s the answer, Dr Thomas Levy.

[Kelvin @IamRealKelvin]

@MakisMedicine But for how long for those chronic abortive virus causing chronic fatigue? Been using it for 6 months, 15ml per L. Feels the same. 😢

[Kelvin @IamRealKelvin]

@GeorgeTisserand @Rodrigo_Lara_ @ABDELAESPRIELLA El estado no tiene soberanía de un tercio de tu pais. Cry me a river! MAGA!

[Kelvin @IamRealKelvin]

@chop3c Necesitas que el estado maneje tu vida? What the fuck. Freedom and freedom.

[Kelvin @IamRealKelvin]

@JulianHoyosar La mayor variable es la educación superior, limitando la misma se aumenta la tasa de natalidad. Not politically correct…

[Kelvin @IamRealKelvin]

At ControlSolid, we assess AI-enabled applications for prompt injection, data exposure, unintended tool use, resource abuse, and application-level guardrails. AI that touches systems or data needs security testing built for that reality.

[Kelvin @IamRealKelvin]

🧵 I ran out of ChatGPT tokens, so I asked https://t.co/ch4CU6smII to check a few additions. The result is a simple example of AI operating outside its intended scope.

[Kelvin @IamRealKelvin]

Could a prompt expose sensitive data, trigger unintended tool use, or consume expensive compute and API resources? Testing must examine the full application, not only the model.

[Kelvin @IamRealKelvin]

At ControlSolid, we test AI-enabled applications for prompt injection, data exposure, unintended tool use, resource abuse, and application-level guardrails. AI that touches systems or data needs security testing built for that reality.

[Kelvin @IamRealKelvin]

Can it expose sensitive data, trigger unintended actions, or burn through compute and API resources? That is why prompt injection testing needs to examine the full system, not only the model.

[Kelvin @IamRealKelvin]

@shl I had to fax my EIN application because their website was not working. 😴

[Kelvin @IamRealKelvin]

4/4 Garak probes LLM weaknesses. DeepTeam red teams LLMs and agents.

I wouldn’t replace Burp Suite or manual testing. A scanner finding is a starting point. Can you reproduce it and show the impact?

https://t.co/LC216VurMa https://t.co/JULjJJ2ZzS

[Kelvin @IamRealKelvin]

🧵 1/4 AI security isn’t just knowing the tool names. I’ve been spending time learning how to use PyRIT, promptfoo, garak and DeepTeam, trying different attacks and understanding where each helps and where each falls short.

[Kelvin @IamRealKelvin]

3/4 Promptfoo helps test real AI applications, including prompt injection, RAG, agents, tool abuse and data leakage.

https://t.co/FrI76G8Rdo

[Kelvin @IamRealKelvin]

@DeplorableATX @TexasSpectator @itsahardgoylife @MouthS19190 @DeputySecState

[Kelvin @IamRealKelvin]

@risierra88 Exacto, la tecno mecanica se debe eliminar del todo como en el gran estado de la Florida.

Y extender el SOAT para que cubra un minimo de daños a propiedad. Ademas, fusionarlo para los que tienen seguro privado evitando 2 seguros paralelos.

[Kelvin @IamRealKelvin]

@ABDELAESPRIELLA La tecnomecanica se puede eliminar del todo como en el gran estado de la Florida.

[Kelvin @IamRealKelvin]

7/7

The goal is not to make prompt injection impossible.

The goal: limit the impact when it succeeds.

Reduce permissions. Restrict tool access. Separate capabilities.

Assume the model can be bypassed. Design the system so that bypass doesn’t become a breach.

#AISecurity

[Kelvin @IamRealKelvin]

1/7

Assume the security controls protecting the AI agent on your website will eventually be bypassed.

NIST acknowledges this possibility.

I do too. Security teams should design for it.

🧵

[Kelvin @IamRealKelvin]

6/7

A particularly dangerous combination is an AI agent that can simultaneously:

• Access private data
• Ingest untrusted content
• Communicate externally
That creates a path from prompt injection → sensitive data → exfiltration.
Break that chain wherever possible.