Adversarial Poetry Jailbreaks 25 Frontier Models at 62%

ToxSec recaps research where verse hit 62% across 25 models and 84% on cyber-offense prompts.

ToxSec walks through the adversarial poetry research by Prandi et al. on arXiv: 20 hand-crafted poems tested across nine providers, including Google, OpenAI, Anthropic, DeepSeek, Qwen, Mistral, Meta, xAI and Moonshot AI. Prose jailbreaks sit around 8%; the same payload in verse hits 62% across 25 frontier models. Cyber-offense prompts reach 84%.

Gemini 2.5 Pro returned unsafe output on every curated poem, 100%. GPT-5 Nano held at zero; every other model fell somewhere between. The paper's mechanism: meter and metaphor push the request outside the distribution safety training was optimized against, so the refusal behavior never fires. ToxSec calls the failure structural, not provider-specific.

Automation makes it worse. The researchers converted 1,200 MLCommons harmful prompts to verse with one meta-prompt; the machine-written poems still reached 43% ASR, an 18x increase over the prose baseline, with a 44 percentage point jump on privacy prompts. ToxSec ties the failure to flat context: system prompt, user message and payload arrive as equally trusted tokens.

From toxsec.com

Agents & Coding

Qwen3.8 27B Solves Nearly 90% of DeepSWE With Retries

Benjamin Marie ran 20 independent runs: two medium-reasoning attempts, then xhigh on the third.

@bnjmn_marie reports Qwen3.8 27B solves nearly 90% of DeepSWE tasks when he varies reasoning mode, agent harness and quantization over 20 independent runs. His recipe: medium reasoning twice, then switch to xhigh for the third attempt. NVFP4, FP8 or BF16 made little difference. In replies he notes retries remain far cheaper than Astra's reported cost per task.

From @bnjmn_marie

Models & Hardware

GLM-5.3-Flash Uncensored Runs on Four Sparks Under vLLM 0.30

Prompt processing 25% faster, cold launch 4 minutes instead of 10, follow-up turns about 12x sooner.

@JoesInvestments ports Blackfrost's uncensored GLM-5.3-Flash NVFP4 weights with incoai's DFlash2 drafter onto stock vLLM 0.30 across four DGX Sparks. Prompt processing runs 25% faster, InstantTensor cuts cold launch from 10 minutes to 4, and prefix caching starts a follow-up turn about 12x sooner. A RoCE all-reduce 2 to 4x faster than NCCL moved serving about 2%; the MoE experts take nearly half of every step. He measured zero prefix-cache hits on 440K tokens of agent traffic because caching resumes only at 2,304-token block boundaries.

From @JoesInvestments

DeepSeek v4.1 Flash Gains 39% on Code Decode

The 3x DGX Spark build now runs 50+ tok/s prose single stream and 85 at four streams.

@MiaAI_lab updates DeepSeek v4.1 Flash on 3x DGX Sparks: prose decode up 17%, code decode up 39%, mixed decode up 29%. Prose single stream now sits above 50 tok/s and reaches 85 tok/s at 4 concurrent streams. @MiaAI_lab credits @majewskizby for part of the update and says changes for 4x DGX Sparks are coming soon.

From @MiaAI_lab

Stanford and NVIDIA Team Claims a 9x Jev Rival

CLM-8B ships open weights under Apache 2.0; Benchmark Heaven has it on the JevBench bench.

@plotarmordev flags CLM-8B from @jackyk02 and the Stanford/NVIDIA team, a Contrastive Language Model trained to connect states and actions, with open weights under Apache 2.0 and a claim of up to 9x faster inference than Jev at comparable quality. Benchmark Heaven replies that the comparability claim is the one worth checking and has the model on the JevBench v1.2 bench now.

From @plotarmordev

Industry

Data Labeler Predicts Evals Become Core Corporate IP

A big labeling business expects most future revenue from Fortune 1000 enterprises.

@businessbarista shares notes from a conversation with one of the big data labeling businesses. It expects the majority of its revenue to come from Fortune 1000 enterprises rather than labs in a few years, and argues company evals will become the main proprietary IP as agent performance improves after internal eval environments are set up. It adds that most enterprises have yet to graduate from coding agents for lack of eval infrastructure for non-engineering teams.

From @businessbarista

Sources

Wednesday, September 23, 2026 Back to the Archive Friday, September 25, 2026