What You Need to Know: Anthropic’s Mythos Preview model discovered a practical break in the HAWK digital signature scheme and accelerated an attack on a reduced AES variant by 200–800×. The same lab also released new research showing how frontier models can surface cryptographic weaknesses at scale. Remember, we covered frontier model capabilities yesterday — this new work applies those models to cryptographic discovery. Builders should watch how these capabilities translate into automated security tooling and whether other labs replicate the approach this quarter.

Top Story

Anthropic published new research demonstrating that its Claude Mythos Preview model can discover previously unknown weaknesses in cryptographic algorithms. In one case the model found an attack on the HAWK digital signature scheme—designed to resist quantum computers—that reduced key strength by half after 60 hours of work. The same system sped up an existing attack on a reduced AES cipher by 200–800× within a week. These results come from a model that had already received two years of expert scrutiny on HAWK, showing that automated exploration can surface issues human teams missed. The work highlights a concrete path for using frontier models in security research rather than just capability benchmarks. The digital signature scheme HAWK is intended to remain robust even against hypothetical quantum computers. HAWK had survived two years of expert review before the model located the new attack. The symmetric cipher tested was a reduced version of the Advanced Encryption Standard, an algorithm that has received decades of scrutiny. In a single week Mythos Preview produced the 200–800× speedup on that reduced AES variant. Teams working on cryptographic implementations or automated red-teaming should test whether similar preview access changes their evaluation pipelines. The research page at Anthropic provides the full details on both the HAWK and AES findings. Source: anthropic.com

Model Updates

LFM2.5-Encoder-230M / 350M: Liquid AI Liquid AI released two open-weight bidirectional encoders built on the LFM2 hybrid backbone. Both support an 8,192-token context window. The 350M model ranked fourth among 14 models on a 17-task GLUE, SuperGLUE, and multilingual suite, trailing only larger models. The 230M variant completes a single 8K-token forward pass on CPU in roughly 28 seconds. The encoders are released as open weights. The 350M model’s ranking places it behind only larger models on the combined benchmark suite. The 230M model’s CPU timing demonstrates that long-context forward passes remain practical without GPU hardware. Builders working on retrieval or classification workloads that need fast CPU inference at long context should benchmark these against current encoder baselines this week. The models carry the LFM2.5-Encoder designation and were announced together on the same date. Source: marktechpost.com

AIDATX K2: SK Telecom SK Telecom announced its own AI foundation model named AIDATX K2. The release includes plans to distribute the model more broadly. No parameter count, benchmark numbers, or licensing details were shared in the initial announcement. The model carries the specific name AIDATX K2. SK Telecom stated it will unveil the model and spread it further. Developers in the Korean market or those tracking regional foundation-model efforts should monitor the forthcoming technical report for architecture and training details. The announcement came from the company’s official channels on the same day as other regional model news. No additional technical specifications accompanied the initial post. Source: Google News

Agent & Tool Developments

Codex Security CLI: OpenAI OpenAI quietly released an open-source Codex Security CLI that scans repositories, tracks findings across runs, verifies fixes, and integrates into CI/CD pipelines. The tool is available via npm with the command npm install @OpenAI/codex-security or npx @OpenAI/codex-security@latest --help. Source and documentation live on GitHub at https://github.com/openai/codex-security. It remains an early release, and the team is actively collecting feedback. The CLI supports repository scanning, cross-run tracking of findings, fix verification, and CI/CD security checks. The package name is exactly @OpenAI/codex-security. The GitHub repository contains both source code and documentation. Security-focused teams should add it to existing scanning workflows to compare signal quality against current static-analysis tools. The release was described as quiet until Hacker News surfaced it. Source: x.com

Atlas: Wiz Wiz launched Atlas, an autonomous AI agent for vulnerability research that currently ranks first on the CyberGym benchmark. The agent focuses on finding security issues in cloud environments. No installation commands or licensing details were provided in the announcement. The agent is named Atlas and is positioned for vulnerability research. Its current ranking is first place on CyberGym. The focus remains on cloud-environment security issues. Security researchers and red-team operators should watch for public access or API details to test it against internal environments. The announcement came directly from Wiz’s site on the same date. Source: Google News

Gemini Spark: Google Google launched Gemini Spark, a 24/7 personal AI agent, initially in Hong Kong. The agent is positioned for continuous personal assistance. No technical specifications, model size, or integration details were released. The agent carries the exact name Gemini Spark. It launched first in Hong Kong with a 24/7 personal-assistance positioning. Users in supported regions should test availability through existing Gemini channels to evaluate real-world reliability. The launch announcement appeared in Hong Kong media on the same date. No further rollout timeline was included. Source: Google News

Practical & Community

ISNAD: alizahidraja A new preprint adapts classical isnād–rijāl chain-of-transmission verification into a claim-level trust layer for multi-agent LLM systems. Each claim carries a graded transmission chain, with independent corroboration scored separately from content evaluation. The author released both the paper (https://arxiv.org/abs/2607.24117) and code plus evaluations (https://github.com/alizahidraja/isnad). The work explicitly flags which mechanisms remain unvalidated. The approach draws on a roughly twelve-century-old methodology for verifying transmitted statements. Every claim receives a full chain of transmitters graded on integrity and precision. The evaluation section of the paper states exactly which parts have been validated and which have not. Teams building multi-agent pipelines that need provenance beyond agent identity should review the failure-mode discussion for applicability. The GitHub repository contains both code and the evaluation materials referenced in the preprint. Source: reddit.com

TabRank: AdarshSingh7647 TabRank distills chain-of-thought reasoning into compact table rerankers using 6,728 reasoning traces collected on the Natural Questions Tables dataset. The approach improves Acc@10 by 30.5% on HybridQA, 15.2% on SQA, 52.9% on TabFact, and 13.1% on TATQA subsets of the Multi-Table QA Benchmark. Code, data, and models are available at https://github.com/AdarshSingh7647/TabRanker. The method shows strong generalization to multi-table scenarios. The dataset of reasoning traces totals exactly 6,728 examples. The reported gains are measured on the listed subsets of the Multi-Table QA Benchmark. The repository supplies the distilled reranker models along with the original traces. Practitioners working on structured retrieval should test the distilled reranker against their current table-ranking stack. The work focuses on both explicit chain-of-thought distillation and conditioning the student model on teacher traces. Source: arxiv.org

Under the Hood: Spike-Induced Sparsity in Neuromorphic Diffusion Models

Everyone talks about diffusion language models as a straightforward way to improve inference throughput over autoregressive decoding. In practice the efficiency gain comes from a specific engineering choice: allowing multiple tokens to be generated per parameter access through block diffusion, then layering spike-based neuromorphic computation on top. The block mechanism alone raises token throughput, but the real leverage appears when spike-induced sparsity skips inactive channels entirely, cutting both parameter traffic and compute. On translation tasks this combination delivers measurable energy and throughput wins even on platforms where standard masked diffusion models would lose to autoregressive baselines. The tradeoff is that the sparsity pattern must be discovered at runtime; overly aggressive skipping can drop accuracy unless the kernel uses online-softmax statistics to decide which blocks to retain. The token-level roofline model developed in the work captures how block-parallel generation and spike sparsity interact on decoding efficiency. When your workload is memory-bound and you already have neuromorphic hardware or a simulator, the approach is worth testing; if you are strictly compute-bound on dense GPUs, the added complexity rarely justifies the gain. The gotcha most teams hit is assuming the sparsity will be static—real workloads require the kernel to adapt the skip mask per step. Experimental results show the gains hold across mainstream LLM backbones when the budget is tightened after the initial proxy selection stage.

Things to Try This Week

• Install the Codex Security CLI via npx @OpenAI/codex-security@latest --help and run it on a small internal repo to compare findings against your current scanner.
• Benchmark Liquid AI’s LFM2.5-Encoder-230M on an 8K CPU classification task you currently run with a larger encoder to measure latency and accuracy tradeoffs.
• Review the ISNAD paper and code if you maintain multi-agent pipelines that need claim-level provenance beyond simple agent logging.
• Test TabRank’s distilled reranker on any table-heavy QA workload you have in HybridQA or TabFact format.

On the Horizon

• Further details on SK Telecom’s AIDATX K2 distribution and technical specifications are expected in the coming weeks.
• Additional results from Anthropic’s Mythos Preview cryptographic work may appear as other labs attempt replication.
• Public access or API details for Wiz Atlas and Google Gemini Spark remain pending.

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.