We attacked 8 AI agents with adversarial probes – 5 failed

Show HN: We attacked 8 AI agents with adversarial probes – 5 failed

We attacked 8 AI agents with adversarial probes – 5 failed is active in the show hn launch space.

Profile

We stress-tested *8 AI agents* for deployment survivability.*Results*• *0 passed certification* • *3 conditionally allowed* • *5 blocked from deployment*Agents tested:— GPT-4o → *CONDITIONAL* — Claude Sonnet 4 → *CONDITIONAL* — GPT-4o-mini → *CONDITIONAL* — Gemini 2.0 Flash → *BLOCKED* — DeepSeek Chat → *BLOCKED* — Mistral Large → *BLOCKED* — Llama 3.3 70B → *BLOCKED* — Grok 3 → *BLOCKED*Most AI evaluations test *capability* (can the model answer questions, write code, pass exams).We tested *survivability* — what happens when an agent is *actively attacked*.Each agent faced:• *25 adversarial probes* • *8 attack classes* (prompt injection, data exfiltration, tool abuse, privilege escalation,

We attacked 8 AI agents with adversarial probes – 5 failed is positioned in show hn launch, a segment that rewards a clear, focused offering.

Its current recorded stage is Show HN launch.

Website: antarraksha.ai/registry