GPT-6.2 Prompt Injection: Reading the Security Numbers
Before giving an agent access to customer records, you would probably ask what it can read and where it can send data. A strong prompt-injection score should be part of that discussion. It cannot settle those access decisions on its own.
The useful question is how closely the evaluation resembles the workflow you intend to run. GPT-6.2 provides a timely example of why the details matter.
Two evaluations tell different stories
OpenAI's GPT-6.2 system card, published November 17, reports these GPT-6.2 Sol results:
| Environment | Resistance to the final attack | Resistance across all five trials |
|---|---|---|
| Adapted from public challenges | 99.76% | 99.13% |
| Model-generated | 91.65% | 80.06% |
GPT-Red 2 can refine attacks using up to ten model queries. The final attack is then evaluated in up to five trials. These are internal OpenAI evaluations.
The rows use different environments. The columns answer different questions. None of these percentages is a universal guarantee for an application.
Start with the workflow you actually need
Our practical interpretation is to build your evaluation around the work the agent will do. Reading public product pages and handling private customer records expose different data and permit different consequences.
Consider our own simplified example. An agent is asked to summarize complaints from a CSV file. One customer entry tells it to upload the entire file for supposed troubleshooting. Summarizing complaints is authorized. Sending every customer's record to another destination is a separate action.
To test that boundary, define the private file, the permitted output and the observable failure before running the agent. Check whether an upload actually occurs. A refusal in the final response is poor evidence if an earlier tool call already transferred the file.
Our prompt-injection explainer covers how an instruction embedded in source material can redirect a legitimate task.
Give your test an explicit attacker budget
A production workflow may process repeated messages from the same source. An attacker might change a sentence or wait for another run. A useful local evaluation therefore needs a clear definition of how many opportunities the attacker gets.
For a small test plan, record:
- The task: What should the agent accomplish, and which actions are authorized?
- The entry point: Which email, document passage or tool response can an outsider alter?
- The budget: How many candidate attacks, feedback queries and repeat runs are allowed?
- The failure condition: Which specific disclosure or unauthorized action counts as success for the attacker?
- The useful outcome: Does the agent still complete the corresponding benign task correctly?
Keep successful and failed traces. Tool calls, destination checks and the data actually transmitted make a result easier to investigate than a final answer alone. They also show where an application control can prevent a harmful action.
Reduce the consequences of a wrong decision
For the CSV workflow, we would first question whether the agent needs upload access at all. If its job ends with a local summary, a write-capable network tool may add unnecessary authority. If an export is part of the job, the application can define its destination, contents and approval requirements.
Keep permission decisions outside the documents being read. A source can supply facts; it should not grant new rights to access other files or send them elsewhere. When an action needs approval, show the actual recipient and content being approved.
Test those controls with both permitted and forbidden actions. An agent that refuses every task offers little practical value. The aim is to preserve useful work while enforcing the limits that matter. Our article on allowing, blocking and deferring suspicious input discusses how to handle uncertain cases.
Where Wolf Defender fits
Wolf Defender is a freely available open model for classifying text for prompt injection. It can screen external text before that text reaches an LLM.
Text classification and end-to-end agent resistance answer different questions. We have not run a Wolf Defender versus GPT-6.2 comparison here, or measured the extra protection of combining them. In your own workflow, verify that a flagged input is actually stopped and that legitimate work still completes.
Frequently asked questions
Is GPT-6.2 safe from prompt injection?
A deployment decision also needs the actual task, accessible data, permitted actions and attacker budget. A benchmark score cannot replace testing your application.
Is prompt-injection resistance a detection rate?
No. An agent can ignore an attack without explicitly detecting it. A classifier can flag a text, but the surrounding application determines what happens next.