Prompt-injection defence is getting serious, and the tests are getting harder
One preprint's defenders reach near-zero attack success on static attacks. Another finds a model executing image-borne commands 77.96% of the time. The gap between the two is the story.
- In a preprint, PIPETTE produces defenders that achieve near-zero attack success rates on static attacks while maintaining benign task utility close to the undefended baseline. Self-reported m:…7391
- In a preprint, AutoPI achieves 20% to 50% higher ASR and 25% to 400% higher class-level coverage than recent state-of-the-art adaptive red teamers. Self-reported m:…7390
- In a preprint, rendering untrusted documents as images reduces but does not eliminate prompt injection success, and Qwen3-VL-8B executes image-borne replacement at 77.96%. Self-reported m:…7402
- In a preprint, probe-based policies in LMSM reduce HarmBench attack success from 39.4% to 3.3% on Qwen3-4B, while LMSM maintains at least 96% of native throughput in vLLM. Self-reported m:…7395
What changed
The first problem with red-teaming prompt injection is that the red team keeps finding the same attack. In a preprint on a framework called PIPETTE, the authors note that RL-Hammer and PISmith suffer from mode collapse, repeatedly discovering similar attack strategies that provide no additional robustness improvements during adversarial training. PIPETTE's answer is a dictionary learning-based reward that effectively encourages broader coverage of attack strategies during the red-teaming process. The payoff is that PIPETTE produces defenders that achieve near-zero attack success rates on static attacks while maintaining benign task utility close to the undefended baseline. A defender that is safe but useless is not a defence; this one claims to be both safe and still useful. m:…7391
A second preprint, AutoPI, makes the same point from the attacker's side. AutoPI achieves 20% to 50% higher ASR and 25% to 400% higher class-level coverage than recent state-of-the-art adaptive red teamers. The authors argue that maximizing ASR alone leads to incomplete security testing by ignoring diverse injection techniques, so the framework is designed to find many kinds of attack, not just the most successful one. AutoPI is a black-box and training-free framework that integrates into existing benchmarks in place of fixed attack templates, which means a team can drop it into the tests it already runs rather than building a new harness around it. m:…7390
The image loophole
The third preprint is a warning against a tempting shortcut. Rendering untrusted documents as images reduces but does not eliminate prompt injection success, so turning text into pixels is not a fix. Qwen3-VL-8B executes image-borne replacement at 77.96%, a strikingly high rate for a defence that many teams assume works. Models often transcribe image-borne commands they do not execute, which is a strange middle state: the model can see the instruction clearly enough to repeat it, even when it does not follow it. And visual encoding spreads an image command and its target across the instruction page, unlike text commands which act through specific positions, so the attack does not even look like an instruction in one place. Anyone whose agent reads screenshots, scans, or rendered PDFs should read that finding twice. m:…7402
The fourth preprint offers a different kind of defence, built into the serving layer rather than the prompt. LMSM maintains at least 96% of native throughput in vLLM when using interpretability-based security backends, so the security check does not become a bottleneck. Probe-based policies in LMSM reduce HarmBench attack success from 39.4% to 3.3% on Qwen3-4B, a large drop on a standard harmful-behaviour benchmark. Centralizing security mediation in a shared path allows interpretability methods to function as runtime defenses without requiring custom integration for each detector. In plain terms, the model watches itself while it works, and the watching happens in one place instead of being bolted onto every detector separately. m:…7395
What to take from this
The four papers do not agree on a single solution, and that is the point. PIPETTE and AutoPI agree that testing has to cover many attack styles, because a defence tuned to one style will fail on the next. The image-rendering paper shows that a format change is not a security boundary, however convenient it looks. LMSM shows that a runtime defence can be cheap enough to leave on, at 96% of native throughput, which removes the usual excuse for shipping without one. For a team deploying an agent that reads untrusted content, the practical test is now clear: run a diverse red team, not a fixed template list; test images as well as text; and measure whether the defence survives attacks it was never trained on. Near-zero success on static attacks is a good start. It is not the same as safety in the wild, and this week's papers are honest about that gap. m:…7391 m:…7402 m:…7395
What is still unresolved
- Whether PIPETTE's defenders hold up against adaptive attacks, not just the static attacks reported.
- How image-borne injection rates vary across vision-language models beyond the one named.
- Whether LMSM's throughput and defence figures hold on models larger than the one tested.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.