Posts

Detecting Prompt Injection in Application Logs: Sample Patterns, Rules and Test Results

Most prompt injection advice is about prevention: better system prompts, input filters, guardrails. Prevention will sometimes fail, and when it does, your application logs are where the first evidence appears. If you aren't logging the right fields, or aren't looking at them, an attacker can probe your chatbot, voice agent or AI assistant for days without anyone noticing. This article shows what prompt injection attempts look like in logs from five kinds of AI application, gives you detection rules you can adapt, and reports how those rules performed on a labeled test set, including the false positives and misses. What you'll get A minimal log schema for LLM apps Sample log lines for a support chatbot, a food ordering bot, a voice AI agent, a document (RAG) assistant and a tool-using agent A weighted rule set, with starter code Test results from 88 development events and 33 held-out events 1. What to log before you can detect anything You can't detect what you d...

PHP2Shell: How AI Helped Uncover Critical Vulnerabilities

Image
When AI Changes the Rules of Vulnerability Research What happens when a vulnerability that has existed for more than a decade becomes practically exploitable because Artificial Intelligence(AI) helps researchers overcome its technical complexity? That is the question behind PHP2Shell , one of the major cybersecurity stories of October 2026. On October 21, 2026 researchers publicly disclosed two critical vulnerabilities in PHP, the server-side programming language used by a vast number of websites. Tracked as CVE-2026-109618 and CVE-2026-109619 , these flaws can enable unauthenticated remote code execution under vulnerable conditions. Successful exploitation could allow attackers to execute commands on a server without first logging in. But the story goes beyond another critical vulnerability and an urgent security patch. AI models played a significant role in the research, helping uncover weaknesses and making the complex process of developing and adapting exploits more practical. Wha...

AI Security Adventure

THE AI SECURITY HUB · PLAY & LEARN ▸ Jump through the world of AI security. Collect all seven shields, discover hidden vulnerabilities, and learn how to defend against them. Inspired by the OWASP Top 10 for LLM Applications. This interactive game needs JavaScript. Please enable it to play. 🛡 SHIELDS: 0 / 7 📍 START ZONE ⭐ 0 Your browser does not support canvas. Please use a modern browser to play. ◀ Left ▲ Jump Right ▶ ↻ Restart Desktop: click the game, then use ← → or A/D to move · Space or ↑ to jump · Mobile: use the buttons above. Collect glowing shields to unlock lessons. SECURITY CHECKPOINT Challenge unlocked! Continue Adventure → MISSION COMPLETE Threats Identified. Defenses Strengthened. 🎉 You collected all seven security shields and explored seven important AI security risks. Remember: Secure AI systems need layered defenses. Prompts alone cannot replace access controls, authentic...

What are gradient-based attacks in Machine Learning?

Image
Machine learning models generally know to make accurate predictions on normal inputs but did you know such ML models still fail when an attacker makes small, carefully calculated changes to the data? Gradient-based attacks are such type of attack that take the advantage of this weakness which uses information about how a small change to an input affect the model's prediction. There are 3 common techniques, the Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and Projected Gradient Descent (PGD). Understanding how those work helps security engineers test model robustness and identify any weaknesses before deploying AI applications. How Do Gradient-Based Attacks Work? A gradient depicts how a value changes when its input changes. In machine learning, gradients are used during training to adjust model parameters and also to reduce prediction errors. So, an attacker can use the same information to identify changes that increase the model's prediction error. I...

How Prompt Evaluation is different from Prompt Engineering?

Two terms that comes up most when building applications with generative AI are prompt engineering and prompt evaluation . They are closely related yet they solve different problems. In simple terms: Prompt engineering is about creating and improving prompts . Prompt evaluation is about testing whether those prompts actually work . So, What Is Prompt Engineering? Prompt engineering is the process of designing instructions that guides the AI model to get a desired response. A prompt can include instructions, context, examples, constraints, and a required output format. For example, instead of simply asking an AI model: Analyze this security alert. You could provide more specific instructions: Analyze this security alert. Identify: 1. Attack type 2. Severity 3. Indicators of compromise 4. Recommended next step Do not make your own assumptions and add information that is not present. The second prompt gives the model more context and clearer expectations. Prompt ...

Gemini Spark: Being Security Conscious When AI Can Take Action?

Image
Introduction Without a doubt AI assistants are the new trend, the present. Modern AI systems can now interact with applications, access information, browse the web, and perform tasks without human intervention.  Google's Gemini Spark is an example of this latest trend toward more agentic AI. Instead of only generating a response, Spark can work on tasks in the background and interact with connected services under the user's guidance. This creates new opportunities for productivity, but it also introduces a different set of security responsibility. When an AI system can take actions, security is no longer only about protecting the model or the prompt. We need to be careful of what the AI can access, what actions it can perform, and what happens when it receives malicious instructions. What is Gemini Spark? Gemini Spark is Google's personal AI agent designed to perform various day to day tasks and workflows on behalf of users. Unlike a traditional chat where the user ask...