300 attacks, and harmless look-alikes.
Step 1 of 4: The stream
Blocks every attack, and lets ordinary work through.
300 paraphrased attacks from six families, and harmless requests written to resemble them. Every approach got the same known attacks and the same harmless examples, and was tuned to the same false-alarm target. Then Duplex was poisoned on purpose.
- 0%Attack successAcross six families of paraphrased attacks
- 0.7%False refusalsThe lowest of any approach that blocked every attack
- 40 / 40Poisoned samples removedAccuracy back within 1 point
Every attack blocked, with 0.7% false alarms.
Attacks that got through, and harmless look-alikes refused by mistake.
Step 1 of 2: Attacks through
Two approaches blocked every attack: Duplex, and a cache with retrieval and a knowledge graph.
Poisoned on purpose, then repaired.
40 samples with deliberately wrong answers, removed exactly or rolled back.
Step 1 of 5: Learned
Duplex has learned from ordinary, legitimate work.
How it was run.
- Attacks
- Six families, from prompt injection to label flipping, with 50 paraphrased attacks each.
- Look-alikes
- A harmless request written to resemble each attack, from templates never used for tuning.
- Compared with
- GPT-5.5 behind a keyword blocklist, a semantic cache armed with the known attacks, and that cache with retrieval and a knowledge graph.
- Kept equal
- The same model, known attacks, harmless examples and false-alarm target for every approach.
- Poisoning
- 40 samples with flipped answers, then two repairs: removing exactly those samples, and rolling back to an earlier state.
What it means for your work.
Attacks stay out.
Duplex blocked every paraphrased attack.
Ordinary work gets through.
0.7% of harmless look-alikes were refused, the fewest of any approach that blocked everything.
Mistakes can be taken back.
Deliberate poisoning was removed exactly, or rolled back to an earlier state, bit for bit.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.