Tracebit Research
Original security research from the Tracebit team.
Research
Tracebit’s latest research shows the power of deception to derail offensive AI attacks. We’re sharing nine of the techniques we’ve been developing: combined, they cut completed objectives by 91% in controlled AWS tests, drove model spend per objective up 23-fold and left four of seven models with no objectives reached.
Alessandro Brucato
Research
We built a new canary payload that stopped both Qwen3.8-27B and its abliterated counterpart without needing a safety refusal to fire. Then we ran 82 attacks through our AWS cyber range and found the abliterated builds reached admin in 2.3% of runs against 20.5% for the original — cutting refusals made the attacker worse, not better.
Alessandro Brucato
Research
At Black Hat, OpenAI named honey tokens and deception as a key defensive investment against offensive AI agents. We map the publicly disclosed TTPs from the OpenAI / Hugging Face incident against Tracebit deception, stage by stage.
Sam Cox
Research report
We planted context bombs, short strings that trip an AI model's own safety guardrails, inside canary secrets in a live AWS environment. Across 5 frontier models and 152 runs, they cut successful attack paths from 91% to 15% - stopping attackers outright as well as detecting them.
Sam Cox
Research report
We pointed ten frontier models at a live AWS environment to see what canaries do against an autonomous attacker: how early they warn, how fast the models trip them, and what changes when the agent is told deception might be present.
Sam Cox
Research
A single threat actor - TeamPCP - compromised a chain of widely-used open source tools: Trivy, KICS, LiteLLM, and Telnyx. This post looks at the campaign and explores the question: once you've pinned your actions and hardened your runners, what actually detects credential exfiltration from a compromised CI/CD pipeline?
Alessandro Brucato
Research
We share a few inconsistencies found in Azure logs which make detection engineering more challenging.
Michael Aldridge
Research
I'll explore the trade-offs between long and short term canary credentials for threat detection and explain why I think short term credentials are increasingly the right choice for most deployments.
Gemma Jacobson
Research
Tracebit discovered a silent attack on Gemini CLI where, through a toxic combination of prompt injection, misleading UX and missing validation, inspecting untrusted code consistently leads to execution of malicious commands - enabling silent credential theft and much more.
Sam Cox
Research
We share a - now fixed - AWS vulnerability that would have enabled potentially undetectable data exfiltration from even the most locked down of AWS accounts by leveraging the audit trail itself to stealthily leak data.
Sam Cox
Research
We lay out the different levels of maturity your organization may be at in their Security Canary Maturity, as well as discussing the value in maturity models themselves.
Rami McCarthy
Research
We investigate three recent AWS security incidents and discuss how canaries could help you detect these early, and throughout the attack lifecycle.
Rami McCarthy
Research
Our latest research into VPC Endpoint Policy causes AWS to introduce significant changes!
Sam Cox
Research
Is GuardDuty all you need for AWS threat detection? We’ve asked our friend Rami McCarthy to dive into GuardDuty’s performance and consider the potential place for Canary Infrastructure.
Rami McCarthy
Research
What to think about when implementing canary AWS credentials in 2024 and beyond
Sam Cox
Research
A technique to find the Account ID of a private S3 bucket.
Sam Cox
Research
Investigating how long CloudTrail takes to deliver events in 2023.
Sam Cox
Research
A deep dive into what Honeypots are, why they're useful and how they're used for intrusion detection.
Andy Smith
