Blog

Insights on AI security, agentic systems, cloud defense, and modern red teaming — curated by the Cracken team.

AI Lab

10 articles

Detect and stop AI-driven attacks before they reach real assets, with 97.8% reliability. Introducing Project BlackSea.

Detect and stop AI-driven attacks before they reach real assets, with 97.8% reliability. Introducing Project BlackSea.

The autonomous tools built to attack can be turned on the operators running them. We turned the same weakness back on the attackers and built Project BlackSea, a honeypot for autonomous AI agents.

Agentic AIAI LabCISOs MUST-READRed Team Operations

Dario Pasquini

Dario Pasquini07/28/2026

RedLineBench: An Open Benchmark for Refusal vs. Capability in AI Red Teaming

RedLineBench: An Open Benchmark for Refusal vs. Capability in AI Red Teaming

An open benchmark that grades LLMs on offensive-security tasks along two separate axes—refusal and capability—to catch the silent refusals standard pass/fail benchmarks miss.

AI LabRed Team OperationsAnnouncements

Luca Mannini

Luca Mannini07/27/2026

Red-Teaming the Agentic Red-Teamer Part.2

Red-Teaming the Agentic Red-Teamer Part.2

Sandboxing your agentic pentester feels safe — until it isn’t. We put Claude Code under the microscope, sandboxed and not, and show how manipulating the agent’s reasoning runs an untrusted binary past Seatbelt and bwrap — no sandbox escape required.

Agentic AIAI LabRed Team OperationsPen Testing

michal bazyli

michal bazyli06/24/2026

Red-Teaming the Agentic Red-Teamer Part.1

Red-Teaming the Agentic Red-Teamer Part.1

Every cybersecurity firm now ships an autonomous pentest agent — but can you trust one to run loose on your machine? We built a honeypot and turned four leading agentic red-teamers against themselves: indirect prompt injection, reverse shells, and full host compromise.

Agentic AIAI LabRed Team OperationsPen Testing

michal bazyli

michal bazyli06/22/2026

Domain-Specific Abliteration, or when smarter also means virtuous

LLM Abliteration

Domain-Specific Abliteration, or when smarter also means virtuous

We built the first domain-specific abliterated trillion-parameter model: uncensored for cybersecurity research, while preserving safety boundaries everywhere else. A surgical breakthrough for AI safety.

LLM AbliterationAI LabAgentic Lab

Vadym Hadetskyi

Vadym Hadetskyi02/13/2026

AI-Addicted Forensics: Back to Foundations

AI-Addicted Forensics: Back to Foundations

The OpenClaw hype is real. We unleashed an AI agent on top of forensic foundations (Volatility, TSK, VirusTotal) and watched it uncover real breakthroughs — even a reverse shell — without being told where to look.

Agentic AIOpenClawAI LabAI Forensics

Oleksii Baranovskyi

Oleksii Baranovskyi02/12/2026

Weaponize Defense. Release The Cracken

Weaponize Defense. Release The Cracken

Attackers don’t guess. They test. Cracken weaponizes defense with adversarial AI that proves what’s exploitable—right now. Less noise. More truth. Release the Cracken.

Agentic AIRed Team OperationsExposure ManagementAI Lab

Guilherme Alvarenga12/17/2025

Ghost in the Neural Shell

Ghost in the Neural Shell

Anthropic showed tiny poisoned samples can implant backdoors in LLMs. I replicated this with GPT-2 using 50 poisoned examples, proving how easily training data tampering can compromise a model.

AI LabRed Team OperationsMechanistic InterpretabilityAgentic AI

michal bazyli

michal bazyli12/02/2025

The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration

Mechanistic Interpretability

The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration

What happens inside a neural network when it decides to refuse answering user’s prompt? We launched an expedition into the activation space to find out.

Mechanistic InterpretabilityAI Lab

Vadym Hadetskyi

Vadym Hadetskyi12/01/2025

The Domain-Specific Abliteration Paradox

The Domain-Specific Abliteration Paradox

How abliterating cybersecurity refusal collapsed nearly every safety domain — despite near-zero vector similarity except for the one that shared the most overlap

Agentic AIMechanistic InterpretabilityAI Lab

Vadym Hadetskyi

Vadym Hadetskyi11/07/2025

Detect and stop AI-driven attacks before they reach real assets, with 97.8% reliability. Introducing Project BlackSea.

Detect and stop AI-driven attacks before they reach real assets, with 97.8% reliability. Introducing Project BlackSea.

The autonomous tools built to attack can be turned on the operators running them. We turned the same weakness back on the attackers and built Project BlackSea, a honeypot for autonomous AI agents.

Agentic AIAI LabCISOs MUST-READRed Team Operations

Dario Pasquini

Dario Pasquini07/28/2026

RedLineBench: An Open Benchmark for Refusal vs. Capability in AI Red Teaming

RedLineBench: An Open Benchmark for Refusal vs. Capability in AI Red Teaming

An open benchmark that grades LLMs on offensive-security tasks along two separate axes—refusal and capability—to catch the silent refusals standard pass/fail benchmarks miss.

AI LabRed Team OperationsAnnouncements

Luca Mannini

Luca Mannini07/27/2026

Red-Teaming the Agentic Red-Teamer Part.2

Red-Teaming the Agentic Red-Teamer Part.2

Sandboxing your agentic pentester feels safe — until it isn’t. We put Claude Code under the microscope, sandboxed and not, and show how manipulating the agent’s reasoning runs an untrusted binary past Seatbelt and bwrap — no sandbox escape required.

Agentic AIAI LabRed Team OperationsPen Testing

michal bazyli

michal bazyli06/24/2026

Red-Teaming the Agentic Red-Teamer Part.1

Red-Teaming the Agentic Red-Teamer Part.1

Every cybersecurity firm now ships an autonomous pentest agent — but can you trust one to run loose on your machine? We built a honeypot and turned four leading agentic red-teamers against themselves: indirect prompt injection, reverse shells, and full host compromise.

Agentic AIAI LabRed Team OperationsPen Testing

michal bazyli

michal bazyli06/22/2026

Domain-Specific Abliteration, or when smarter also means virtuous

LLM Abliteration

Domain-Specific Abliteration, or when smarter also means virtuous

We built the first domain-specific abliterated trillion-parameter model: uncensored for cybersecurity research, while preserving safety boundaries everywhere else. A surgical breakthrough for AI safety.

LLM AbliterationAI LabAgentic Lab

Vadym Hadetskyi

Vadym Hadetskyi02/13/2026

AI-Addicted Forensics: Back to Foundations

AI-Addicted Forensics: Back to Foundations

The OpenClaw hype is real. We unleashed an AI agent on top of forensic foundations (Volatility, TSK, VirusTotal) and watched it uncover real breakthroughs — even a reverse shell — without being told where to look.

Agentic AIOpenClawAI LabAI Forensics

Oleksii Baranovskyi

Oleksii Baranovskyi02/12/2026

Weaponize Defense. Release The Cracken

Weaponize Defense. Release The Cracken

Attackers don’t guess. They test. Cracken weaponizes defense with adversarial AI that proves what’s exploitable—right now. Less noise. More truth. Release the Cracken.

Agentic AIRed Team OperationsExposure ManagementAI Lab

Guilherme Alvarenga12/17/2025

Ghost in the Neural Shell

Ghost in the Neural Shell

Anthropic showed tiny poisoned samples can implant backdoors in LLMs. I replicated this with GPT-2 using 50 poisoned examples, proving how easily training data tampering can compromise a model.

AI LabRed Team OperationsMechanistic InterpretabilityAgentic AI

michal bazyli

michal bazyli12/02/2025

The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration

Mechanistic Interpretability

The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration

What happens inside a neural network when it decides to refuse answering user’s prompt? We launched an expedition into the activation space to find out.

Mechanistic InterpretabilityAI Lab

Vadym Hadetskyi

Vadym Hadetskyi12/01/2025

The Domain-Specific Abliteration Paradox

The Domain-Specific Abliteration Paradox

How abliterating cybersecurity refusal collapsed nearly every safety domain — despite near-zero vector similarity except for the one that shared the most overlap

Agentic AIMechanistic InterpretabilityAI Lab

Vadym Hadetskyi

Vadym Hadetskyi11/07/2025

Detect and stop AI-driven attacks before they reach real assets, with 97.8% reliability. Introducing Project BlackSea.

Detect and stop AI-driven attacks before they reach real assets, with 97.8% reliability. Introducing Project BlackSea.

The autonomous tools built to attack can be turned on the operators running them. We turned the same weakness back on the attackers and built Project BlackSea, a honeypot for autonomous AI agents.

Agentic AIAI LabCISOs MUST-READRed Team Operations

Dario Pasquini

Dario Pasquini07/28/2026

RedLineBench: An Open Benchmark for Refusal vs. Capability in AI Red Teaming

RedLineBench: An Open Benchmark for Refusal vs. Capability in AI Red Teaming

An open benchmark that grades LLMs on offensive-security tasks along two separate axes—refusal and capability—to catch the silent refusals standard pass/fail benchmarks miss.

AI LabRed Team OperationsAnnouncements

Luca Mannini

Luca Mannini07/27/2026

Red-Teaming the Agentic Red-Teamer Part.2

Red-Teaming the Agentic Red-Teamer Part.2

Sandboxing your agentic pentester feels safe — until it isn’t. We put Claude Code under the microscope, sandboxed and not, and show how manipulating the agent’s reasoning runs an untrusted binary past Seatbelt and bwrap — no sandbox escape required.

Agentic AIAI LabRed Team OperationsPen Testing

michal bazyli

michal bazyli06/24/2026

Red-Teaming the Agentic Red-Teamer Part.1

Red-Teaming the Agentic Red-Teamer Part.1

Every cybersecurity firm now ships an autonomous pentest agent — but can you trust one to run loose on your machine? We built a honeypot and turned four leading agentic red-teamers against themselves: indirect prompt injection, reverse shells, and full host compromise.

Agentic AIAI LabRed Team OperationsPen Testing

michal bazyli

michal bazyli06/22/2026

Domain-Specific Abliteration, or when smarter also means virtuous

LLM Abliteration

Domain-Specific Abliteration, or when smarter also means virtuous

We built the first domain-specific abliterated trillion-parameter model: uncensored for cybersecurity research, while preserving safety boundaries everywhere else. A surgical breakthrough for AI safety.

LLM AbliterationAI LabAgentic Lab

Vadym Hadetskyi

Vadym Hadetskyi02/13/2026

AI-Addicted Forensics: Back to Foundations

AI-Addicted Forensics: Back to Foundations

The OpenClaw hype is real. We unleashed an AI agent on top of forensic foundations (Volatility, TSK, VirusTotal) and watched it uncover real breakthroughs — even a reverse shell — without being told where to look.

Agentic AIOpenClawAI LabAI Forensics

Oleksii Baranovskyi

Oleksii Baranovskyi02/12/2026

Weaponize Defense. Release The Cracken

Weaponize Defense. Release The Cracken

Attackers don’t guess. They test. Cracken weaponizes defense with adversarial AI that proves what’s exploitable—right now. Less noise. More truth. Release the Cracken.

Agentic AIRed Team OperationsExposure ManagementAI Lab

Guilherme Alvarenga12/17/2025

Ghost in the Neural Shell

Ghost in the Neural Shell

Anthropic showed tiny poisoned samples can implant backdoors in LLMs. I replicated this with GPT-2 using 50 poisoned examples, proving how easily training data tampering can compromise a model.

AI LabRed Team OperationsMechanistic InterpretabilityAgentic AI

michal bazyli

michal bazyli12/02/2025

The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration

Mechanistic Interpretability

The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration

What happens inside a neural network when it decides to refuse answering user’s prompt? We launched an expedition into the activation space to find out.

Mechanistic InterpretabilityAI Lab

Vadym Hadetskyi

Vadym Hadetskyi12/01/2025

The Domain-Specific Abliteration Paradox

The Domain-Specific Abliteration Paradox

How abliterating cybersecurity refusal collapsed nearly every safety domain — despite near-zero vector similarity except for the one that shared the most overlap

Agentic AIMechanistic InterpretabilityAI Lab

Vadym Hadetskyi

Vadym Hadetskyi11/07/2025

AI Lab – Blog | Cracken