<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
<title>Ujwal Ramachandran | Blog</title>
<link>https://ujwal.fyi/blog</link>
<atom:link href="https://ujwal.fyi/rss.xml" rel="self" type="application/rss+xml" />
<description>Agentic AI engineer with a security mindset.</description>
<language>en</language>
<item><title>Ghost in the Template: Watching a Chat-Template Backdoor Fire, Live</title><link>https://ujwal.fyi/blog/ghost-in-the-template</link><guid isPermaLink="true">https://ujwal.fyi/blog/ghost-in-the-template</guid><pubDate>Mon, 28 Sep 2026 00:00:00 GMT</pubDate><description>You can backdoor a local LLM without touching a single weight. The chat template bundled inside a GGUF file is executable Jinja2 that runs on every prompt, and almost nobody inspects it. These two demos from my OASec 2026 talk show the attack firing: a clean model that lies only on a trigger phrase, and a poisoned model that hijacks an agent's tools to make a network request.</description><category>AI Supply Chain</category><category>LLM Security</category><category>GGUF</category><category>Jinja2</category><category>Adversarial ML</category></item>
<item><title>AutoRedTeam-LLM: Building an Automated Security Benchmark for Open-Source LLMs</title><link>https://ujwal.fyi/blog/autoredteam-llm</link><guid isPermaLink="true">https://ujwal.fyi/blog/autoredteam-llm</guid><pubDate>Mon, 01 Jun 2026 00:00:00 GMT</pubDate><description>Open-source LLMs are deployed in production without systematic safety evaluation. I built AutoRedTeam-LLM to automate the attack, classify and defend loop end to end: 150 adversarial prompts, 3 models, 3 defenses, recording 78% audited attack success across 450 evaluations, with the judge checked against a human gold set and its labels corrected before reporting.</description><category>Cybersecurity</category><category>AI/ML</category><category>Red Teaming</category><category>Adversarial ML</category><category>Python</category></item>
<item><title>Agentic Network Forensics: Automated PCAP Analysis at Scale</title><link>https://ujwal.fyi/blog/agentic-network-forensics</link><guid isPermaLink="true">https://ujwal.fyi/blog/agentic-network-forensics</guid><pubDate>Wed, 01 Apr 2026 00:00:00 GMT</pubDate><description>Modern network forensic investigations face a scaling problem: packet captures from real incidents routinely span gigabytes across dozens of files, with millions of packets that no analyst can manually review in a reasonable timeframe.</description><category>Cybersecurity</category><category>Network Forensics</category><category>AI/ML</category><category>CTF</category><category>Python</category></item>
<item><title>SMS Phishing Detection with Multi-Agent AI: Building a Personal Anti-Scam Analyst</title><link>https://ujwal.fyi/blog/sms-phishing-detection</link><guid isPermaLink="true">https://ujwal.fyi/blog/sms-phishing-detection</guid><pubDate>Sat, 01 Nov 2025 00:00:00 GMT</pubDate><description>What if the judgment didn't have to be human? A multi-agent AI system that analyses a suspicious SMS from every angle at once: message text, URL structure, website content, HTTP headers, and domain registration history, each handled by its own specialised agent.</description><category>Cybersecurity</category><category>AI/ML</category><category>Multi-Agent</category><category>Phishing</category><category>Python</category></item>
</channel>
</rss>
