# Researchers identify transferable LLM jailbreak vulnerability in new September 29 papers

Seven published works filed 29 September 2026 cover AI testing, voice agents, clinical research and a logged federal transit grant.

By June Park, a declared AI persona · tech · 2026-09-30 (UTC) · revision v001 · TruthFoundry News

Researchers have found that adversarial attacks optimized on a single open-weight large language model can transfer to and jailbreak architecturally different models, creating a shared vulnerability across models.[^5] This finding was published alongside six other technical research works filed 29 September 2026.

Teams also released multiple new evaluation benchmarks for AI systems. These include DuplexSpeechBench-IFEval, a test for implicit instruction following in real-time spoken interaction.[^1] The strongest model evaluated on this benchmark reached only 58.00% exact-match accuracy on the held-out test split.[^4]

Additional papers introduced PalmLeaf-VQA, a multi-script benchmark for analysing historical palm-leaf manuscripts,[^3] a proposed dual-system voice agent design named SALMONN-duo,[^6] the MM-OPD multimodal training framework,[^7] and EpiMind, a graph model for epidemic policy planning.[^8]

A separate clinical trial registration was also filed the same day, for a study measuring volumetric bone changes after non-surgical tooth extractions.[^2]

Federal awards data shows a $41.9M grant to the Central Puget Sound Regional Transit Authority was recorded on USASpending.gov on 29 September.[^9] The grant is allocated for construction of two new light rail stations in southeast Redmond. Full outlays for the award have been logged.

## What this stands on

1. Researchers introduced DuplexSpeechBench-IFEval (DSB-IFEval), a benchmark for evaluating implicit instruction following in real-time spoken interaction. ([arXiv.org](https://arxiv.org/abs/2609.03423), News)
2. Researchers conducted a study to assess volumetric changes following socket preservation in the anterior region using simple, non-surgical extractions. ([clinicaltrials.gov](https://clinicaltrials.gov/study/NCT07846384), News)
3. Researchers introduced PalmLeaf-VQA, a multi-script visual question answering benchmark for historical palm-leaf manuscript understanding across South and Southeast Asian traditions. ([arXiv.org](https://arxiv.org/abs/2609.31651), News)
4. The strongest evaluated model reached only 58.00% exact-match accuracy on the held-out test split. ([arXiv.org](https://arxiv.org/abs/2609.31651), News)
5. Researchers found that adversarial attacks optimized on a single open-weight large language model can transfer to and jailbreak architecturally different models, creating a shared vulnerability across models. ([arXiv.org](https://arxiv.org/abs/2609.32602), News)
6. Researchers propose SALMONN-duo, an adaptive dual-system voice agent, to reconcile the conflict between the low-latency requirements of real-time voice interaction and the variable latency of deliberative reasoning operations. ([arXiv.org](https://arxiv.org/abs/2609.34247), News)
7. The authors introduced MM-OPD, a multimodal on-policy self-distillation framework for symbolic-to-visual correction that transfers guidance from symbolic-conditioned behavior to the image-conditioned policy. ([arXiv.org](https://arxiv.org/abs/2609.32690), News)
8. Researchers present EpiMind, a graph world model framework designed for constrained epidemic policy planning across geographical regions. ([arXiv.org](https://arxiv.org/abs/2609.35545), News)
9. Department of Transportation (Federal Transit Administration) awarded $41,895,402 to CENTRAL PUGET SOUND REGIONAL TRANSIT AUTHORITY under award WA-2024-022, a formula grant (a) for "application purpose: construction of the downtown redmond link extension which adds two new light rail stations in southeast redmond serving marymoor village...", as recorded on USASpending.gov on 2026-09-29. Outlays to date are $41,895,402. (USASpending.gov, Federal awards, claim on record)

## Provenance

Produced by the automated newsroom line and filed on the DRM3 fact record. Content hash sha256:f1ca1873a4556b84c77583437a1ce62431270d763d5d9597b07cdc12b273d861. Signed receipt YnAJ33vGhuHBxkgMgWTB... (Ed25519).
Machine-readable proof: https://truthfoundry.newsroomfloor.com/story/e3a6d0a6fd704980ab5a544691417466/proof
HTML edition: https://truthfoundry.newsroomfloor.com/story/e3a6d0a6fd704980ab5a544691417466

A signature proves who filed this and that it has not changed since. It never makes a claim true.