Now that DARPA’s AI Cyber Challenge (AIxCC) has officially ended, we can finally make Buttercup, our CRS (Cyber Reasoning System), open source!

While the AIxCC winner has not yet been announced, differences in the finalists’ approaches show that there are multiple viable paths forward to using AI for vulnerability detection.

Prompt injection pervades discussions about security for LLMs and AI agents. But there is little public information on how to write powerful, discreet, and reliable prompt injection exploits. In this post, we will design and implement a prompt injection exploit targeting GitHub’s Copilot Agent, with a focus on maximizing reliability and minimizing the odds of detection.

In my first month at Trail of Bits as an AI/ML security engineer, I found two remotely accessible memory corruption bugs in NVIDIA’s Triton Inference Server during a routine onboarding practice.

Trail of Bits founder Dan Guido establishes a $2,500 scholarship at his alma mater, Mineola High School, to recognize students who demonstrate the hacker spirit through self-driven learning, creative problem-solving, and unconventional technological exploration. The scholarship celebrates tomorrow’s security innovators who push boundaries and think differently about technology.

We’re releasing pajaMAS: a curated set of MAS hijacking demos that illustrate important principles of MAS security.

Today we’re announcing the beta release of mcp-context-protector, a security wrapper for LLM apps using the Model Context Protocol (MCP). It defends against the line jumping attacks documented earlier in this blog series, such as prompt injection via tool descriptions and ANSI terminal escape codes.

We successfully exploited two discontinued network devices at DistrictCon’s inaugural Junkyard competition in February, winning runner-up for Most Innovative Exploitation Technique. Our exploit chains demonstrate why end-of-life hardware poses persistent security risks.

At EthCC[8], Trail of Bits blockchain security engineer Nicolas Donboly laid out a clear, actionable path for aspiring smart contract auditors, drawing from his own experience transitioning from a non-technical background into a leading security role.

Vendetect is our new open-source tool for detecting copied and vendored code between repositories. It uses semantic fingerprinting to identify similar code even when variable names change or comments disappear. More importantly, unlike academic plagiarism detectors, it understands version control history, helping you trace vendored code back to its exact source commit.

The release of Bitchat last week was met with a mixture of glowing praise and sharp criticism. Both extremes bear some truth, but they also miss the mark and reveal gaps in how we discuss security in emerging products.

Deptective, our new open-source tool, automatically finds the packages needed to install software dependencies. It does so not based on the software’s self-reported requirements, but by observing what the software needs at runtime.

Our CRS (Cyber Reasoning System), Buttercup, is now competing in the one and only scored round of DARPA’s AI Cyber Challenge (AIxCC) against six other teams to see which autonomous AI-driven system can find and patch the most software vulnerabilities.

Private key compromise accounted for 43.8% of crypto hacks in 2024, yet traditional smart contract audits rarely address architectural access control weaknesses. This post introduces a four-level maturity framework for designing protocols that can tolerate key compromise, progressing from single EOA control to radical immutability, with practical examples demonstrating multisigs, timelocks, and the principle of least privilege.

File parsers in Go contain unexpected behaviors that can lead to serious security vulnerabilities. This post examines how JSON, XML, and YAML parsers in Go handle edge cases in ways that have repeatedly resulted in high-impact security issues in production systems. We explore three real-world attack scenarios: marshaling/unmarshaling unexpected data, exploiting parser differentials, and leveraging data format confusion. Through examples, we demonstrate how attackers can bypass authentication, circumvent authorization controls, and exfiltrate sensitive data by exploiting these parser behaviors.