Testing and defence led by a published security researcher with five CVEs, including a CVSS 9.1 Mirai botnet kill-switch.
All security →SOC 2, ISO, and the Canadian privacy stack, run end to end with an independent auditor.
All frameworks →Original research, free tools, and plain-language guides on security and compliance, from a published security researcher.
Read the blog →Prompt injection is an attack against AI applications in which crafted input causes a large language model to ignore its original instructions and follow the attacker's instead. Because LLMs cannot reliably separate trusted instructions from untrusted data, attacker-controlled text can hijack the model's behavior. It is the number one risk in the OWASP Top 10 for LLM Applications.
Direct prompt injection comes from the user typing malicious instructions. Indirect prompt injection is more dangerous: the payload hides in content the model later reads, such as a web page, document, or email, and triggers when the model processes it.
In agentic systems the stakes rise, because a hijacked model can call tools, send data, or take actions on the attacker's behalf. There is no single fix; defenses layer input handling, output constraints, least-privilege tool access, and adversarial testing.
traztech delivers prompt injection and LLM adversarial testing for startups and growth-stage companies, led by a published security researcher.
Book a callFor a broader look at getting audit-ready, see our SOC 2 readiness work, or talk to a fractional CISO about building a program around it.
Prompt injection stops being theoretical the moment a model can act: call a tool, read a document, or return output that another system trusts. Until then it is a content problem; after it, it is an authorisation problem.
The mitigations that hold are architectural. Treat model output as untrusted input, constrain what the model can reach, and put authorisation checks between the model and anything consequential.
No. System prompts can be overridden by content the model reads. Defences that work sit around the model rather than inside the prompt.
Where the malicious instruction arrives through content the model retrieves rather than from the user, such as a web page or a document in a retrieval pipeline. It is the harder case because the user never sees it.
Track record
Five published CVEs. CVE-2024-45163 (CVSS 9.1) is a flaw in the Mirai botnet itself, which gave defenders a way to shut down attacker infrastructure. CVE-2026-42626 takes HP ENVY 5000 printers offline from any unauthenticated device on the same network.
At Humera, a venture-backed US security company, Jacob built the compliance programme in-house from nothing: no report, no policies, no documented controls. It ended in a Type II attestation across 76 controls with zero exceptions, on a team of 15.
Before you go
Definitions only get you so far. I send a few short notes on how this plays out in practice. Unsubscribe in one click, and replies reach me directly.
From Jacob Masse, principal of traztech. No spam, unsubscribe in one click.