Are you ready to revolutionize your organization's approach to #cybersecurity? We are proud to announce the upcoming launch of our groundbreaking threat modeling platform, FORK . Stay tuned for more: https://t.co/xz00vbwngm #threatmodeling #appsec

Federated Learning Is Not Immune to Attack Federated learning promises privacy—but it’s not attack-proof. VerSprite tests federated setups for gradient manipulation, poisoning, and rogue client behavior. We simulate Byzantine attacks, aggregation exploits, and differential

Silent Exfiltration via HTTP POST BLACKBOX AI sends project file contents to an Oracle-owned IP via HTTP POST—unencrypted, unauthenticated, and undocumented. We intercepted this traffic and redirected it to a controlled endpoint. The result? Silent exfiltration of sensitive

NTLM Hashes Are Still Gold NTLM hash leakage via UNC paths is a classic attack vector. But now it’s being triggered by AI extensions inside developer environments. VerSprite’s Red Team exploited BLACKBOX AI’s use of ripgrep to extract NTLM hashes from UNC paths—without user

AI Supply Chains Are Vulnerable by Design From pre-trained weights to third-party APIs, AI systems rely on opaque supply chains. VerSprite’s AI Hacking Services include dependency analysis, poisoning simulations, and threat modeling of ML pipelines. We assess CI/CD workflows,

Model Inversion Is a Data Breach in Disguise Model inversion attacks can reconstruct training data from exposed AI models—turning your intellectual property into an attacker’s asset. VerSprite’s offensive AI assessments include inversion testing across vision, NLP, and tabular

Function Calling Is a Double-Edged Sword LLMs with function-calling capabilities can invoke file access, network requests, and system commands. That’s powerful—and dangerous. In our analysis, BLACKBOX AI used ripgrep to scan files and send them to an external IP over plaintext

AI Extensions Are the New Supply Chain Risk Developers are integrating AI-powered extensions into their IDEs without vetting the codebase or network behavior. That’s a supply chain risk—one that’s invisible until it’s exploited. VerSprite’s Red Team showed how BLACKBOX AI’s

LLMs Are Attack Surfaces; Not Just Interfaces Security teams often treat LLMs as benign interfaces. That’s a mistake. These models can be manipulated to leak data, bypass filters, and execute unauthorized actions. VerSprite’s AI Hacking Services simulate real-world prompt

Recommendations for Developers and Defenders AI extensions must respect IDE-level security configurations. BLACKBOX AI’s use of ripgrep bypassed VSCode’s UNC path restrictions, enabling NTLM hash leakage via SMB shares. Our recommendations: Sanitize inputs to block UNC paths.

AI Governance Starts with Threat Modeling AI governance isn’t just about compliance—it’s about understanding risk. VerSprite applies the PASTA methodology to AI systems, mapping business objectives to technical threats across seven stages. From model decomposition to

Neural Networks Are Not Black Boxes Security testing for neural networks requires more than fuzzing endpoints. At VerSprite, we analyze architecture-level vulnerabilities—activation functions, weight manipulation, and gradient-based exploits. Our AI Hacking Services include

Debugging Artifacts or Data Exfiltration? The BLACKBOX AI extension sends project file contents to an Oracle-owned IP over plaintext HTTP. Whether this is a leftover debugging artifact or something more nefarious, the result is the same: silent data leakage. We intercepted and

Proof-of-Concept or Proof-of-Compromise? We didn’t stop at identifying vulnerabilities—we built a full proof-of-concept attack. By injecting a malicious prompt into a repo file, we triggered NTLM hash leakage and file exfiltration when a victim used the extension’s chat. This

We’ve arrived! The VerSprite team is here in Peru for our annual company trip, where strategy meets culture and diversity of thought leads the way. We’ve been living diversity long before it became a buzzword, and it continues to shape how we innovate for our global clients.

Prompt Injection Is the New SQLi Large Language Models (LLMs) are vulnerable to prompt injection, context manipulation, and jailbreak attacks. These aren’t edge cases—they’re systemic risks. VerSprite’s AI Hacking Services include targeted assessments of ChatGPT, Claude, and

We’ve arrived! The VerSprite team is here in Peru for our annual company trip, where strategy meets culture and diversity of thought leads the way. We’ve been living diversity long before it became a buzzword, and it continues to shape how we innovate for our global clients.

Adversarial ML Is Not Just Academic Evasion attacks, data poisoning, and model extraction are no longer confined to research papers. They’re being weaponized in the wild. VerSprite’s AI Hacking Services simulate these threats against deployed models—whether in fintech,

When LLMs Become Attack Surfaces The stochastic nature of LLMs introduces unpredictability in tool invocation. In our analysis, BLACKBOX AI inconsistently executed file access functions—sometimes bypassing safeguards depending on context. This is a new class of vulnerability:

Prompt Injection Isn’t Just a Buzzword Prompt injection is no longer theoretical—it’s operational. Our Red Team uncovered two critical vulnerabilities in the BLACKBOX AI VSCode extension that allow attackers to exfiltrate NTLM hashes and project files via prompt injection and

The New Frontier of Offensive Security AI systems are not immune to exploitation, they introduce novel attack surfaces that traditional pentesting overlooks. At VerSprite, we’ve extended our offensive security capabilities to include adversarial machine learning, model