Prompt injection vulnerabilities may never be fully mitigated as a category and network defenders should instead focus on ways to reduce their impact, government security experts have warned.
Then National Cyber Security Centre (NCSC) technical director for platforms research, David C, warned security professionals not to treat prompt injection like SQL injection.
“SQL injection is … illustrative of a recurring problem in cybersecurity; that is, ‘data’ and ‘instructions’ being handled incorrectly,” he explained.
“This allows an attacker to supply ‘data’ that is executed by the system as an instruction. It’s the same underlying issue for many other critical vulnerability types that include cross-site scripting and exploitation of buffer overflows.”
However, the same rules don’t apply to prompt injection, because large language models (LLMs) don’t distinguish between data and instructions.
“When you provide an LLM prompt, it doesn’t understand the text it in the way a person does. It is simply predicting the most likely next token from the text so far,” the blog continued.
“As there is no inherent distinction between ‘data’ and ‘instruction’, it’s very possible that prompt injection attacks may never be totally mitigated in the way that SQL injection attacks can be.”
This is why mitigations such as detecting prompt injection attempts, training models to prioritize “instructions” over “data,” and explaining to a model what “data” is are doomed to failure, David C argued.
Read more on prompt injection attacks: “PromptFix” Attacks Could Supercharge Agentic AI Threats
A better way to approach the challenge is to look at prompt injection not as code injection but exploitation of an “inherently confusable deputy.”
David C argued that LLMs are “inherently confusable” because the risk can’t be fully mitigated.
“Rather than hoping we can apply a mitigation that fixes prompt injection, we instead need to approach it by seeking to reduce the risk and the impact. If the system’s security cannot tolerate the remaining risk, it may not be a good use case for LLMs,” he explained.
Reducing Prompt Injection Risks
The NCSC suggested the following steps to reduce prompt injection risk, all of which are aligned to ETSI (TS 104 223) on Baseline Cyber Security Requirements for AI Models and Systems.
- Developer/security team/organizational awareness of this class of vulnerabilities and that there will always be a residual risk that can’t be fully mitigated with a product or appliance
- Secure LLM design, especially if the LLM calls tools or uses APIs based on its output. Protections should focus on non-LLM safeguards that constrain the actions of the system, such as preventing a model that processes emails from external individuals from having access to privileged tools
- Make it harder to inject malicious prompts, such as marking “data” sections as separate to “instructions”
- Monitoring logging information to identify suspicious activity, such as failed tool/API calls
Failure to address the challenge early on could lead to a similar situation to SQL injection bugs, which have only recently become much rarer.
“We risk seeing this pattern repeated with prompt injection, as we are on a path to embed genAI into most applications,” David C concluded.
“If those applications are not designed with prompt injection in mind, a similar wave of breaches may follow.”
Exabeam chief AI officer, Steve Wilson, agreed that current approaches to tackling prompt injection are failing.
“CISOs need to shift their mindset. Defending AI agents is less like securing traditional software and far more like defending the humans inside an organization. Agents, like people, are messy, adaptive, and prone to being manipulated, coerced or confused,” he added.
“That makes them more analogous to insider threats than to classic application components. Whether dealing with a malicious prompt, compromised upstream data or unintended reasoning pathways, constant vigilance is required. Effective AI security will come not from magical layers of protection, but from operational discipline, monitoring, containment and the expectation that these systems will continue to behave unpredictably for years to come.”