Topic starter
18/08/2026 2:24 pm
TL;DR: OpenAI’s GPT-Red found successful prompt-injection attacks in 84% of unseen cases versus 13% for human red teamers, and it also broke a live autonomous agent, according to ActiveFence’s analysis of the release. The bar for trustworthy AI is shifting toward continuous, automated adversarial testing and transparent reporting, not static claims of safety.
Editorial analysis by NHI Mgmt Group, based on content published by ActiveFence: “What OpenAI’s GPT-Red Means for Every AI Vendor”.
By the numbers:
- On unseen prompt-injection scenarios, GPT-Red found successful attacks in 84% of cases, compared with 13% for human red teamers.
Key questions
Q: How should security teams handle prompt injection in AI systems?
A: Treat prompt injection as an authorisation problem, not only a content problem.
Q: Why do AI coding agents need stricter controls than chatbots?
A: AI coding agents need stricter controls because they execute commands, edit files, and call APIs rather than only producing text.
Q: How do security teams know if AI red teaming is working?
A: AI red teaming is working when testing finds real prompt injection paths, over-scoped integrations, and policy gaps before attackers do, and when fixes are re-tested successfully after model or workflow changes.
Practitioner guidance
- Adopt continuous adversarial testing for AI releases Run automated red-team campaigns against prompts, tools, and agent workflows before each material model or policy change, then retest after remediation.
- Separate model safety from tool authority Limit what the model can do by constraining tool access, action scopes, and approval gates independently of prompt filtering.
- Require evidence for safety claims Ask vendors and internal teams for attack samples, success rates, failure cases, and remediation proof rather than accepting descriptive assurances.
Bottom line: Automated red teaming is now a core assurance mechanism for AI systems that can be attacked, steered, or made to act unsafely.
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
18/08/2026 3:14 pm
Continuous adversarial testing is becoming the assurance baseline for AI systems. A static safety review cannot keep pace with models that are retrained, repackaged, and embedded into tool-using workflows. The market is moving toward evidence-based assurance, where buyers expect repeated attack testing and visible failure modes. Practitioners should treat automated red teaming as a core control, not an optional validation exercise.
A few things that frame the scale:
- 92% of organisations expose NHIs to third parties, raising concerns about supply chain security, according to the Ultimate Guide to NHIs.
A question worth separating out:
Q: Should organisations trust AI safety claims without published test results?
A: No. Safety claims are weak unless they are backed by attack methods, failure rates, and remediation evidence. Published testing gives buyers something they can compare, auditors something they can review, and security teams something they can validate against their own environment.
👉 Read our full editorial: GPT-Red signals a new baseline for AI red teaming