AI Agents Get Safer, Drug Discovery Accelerates

🛡️ Agents

DTap: Red-Teaming AI Agents

The Rundown

AI agents are increasingly handling high-stakes tasks, raising significant security concerns. Adversaries can exploit these agents to leak sensitive information or perform unauthorized actions. Evaluating agent security in dynamic, real-world environments remains a challenge. To address this, researchers developed the DecodingTrust-Agent Platform (DTap). This is the first controllable and interactive red-teaming platform designed specifically for AI agents. DTap spans 14 real-world domains and over 50 simulation environments, replicating systems like Google Workspace, PayPal, and Slack. It aims to provide a scalable environment for risk assessment. The platform also introduces DTap-Red, an autonomous red-teaming agent. DTap-Red systematically explores various injection vectors—including prompt, tool, skill, and environment manipulations—to discover effective attack strategies. This autonomous approach helps curate DTap-Bench, a large dataset of high-quality red-teaming instances with verifiable judges for automatic outcome validation. Large-scale evaluations using DTap reveal systematic vulnerability patterns in popular AI agents, offering insights for building more secure next-generation systems.

The details

• DTap offers red-teaming capabilities across 14 real-world domains and over 50 simulation environments.
• DTap-Red autonomously explores prompt, tool, skill, and environment injection vectors to find vulnerabilities.
• DTap-Bench includes high-quality red-teaming instances with verifiable judges for automatic attack validation.
• Evaluations using DTap reveal systematic vulnerability patterns in popular AI agents.

Why it matters

Startups building AI agents need robust security testing. DTap provides a structured framework for identifying and mitigating vulnerabilities before deployment, crucial for maintaining user trust and preventing costly breaches.

🚨 Agents

AgentTrust: Runtime Safety for Agents

The Rundown

AI agents execute real-world actions via tool calls, posing risks of irreversible harm from a single unsafe action. Existing defenses fall short—benchmarks test post-execution, static guardrails miss obfuscation, and sandboxes lack action understanding. AgentTrust introduces a runtime safety layer that intercepts agent tool calls before execution. It provides a structured verdict: allow, warn, block, or review. AgentTrust combines a shell deobfuscation normalizer, SafeFix suggestions for safer alternatives, RiskChain detection for multi-step attacks, and a cache-aware LLM-as-Judge for ambiguous inputs. The system was evaluated on a 300-scenario benchmark across six risk categories and an additional 630 adversarial scenarios. Under a patched ruleset, AgentTrust achieved 96.7% verdict accuracy, including about 93% on shell-obfuscated payloads, with low-millisecond latency. The platform is released under the AGPL-3.0 license and supports Model Context Protocol-compatible agents.

The details

• AgentTrust intercepts tool calls at runtime, providing allow, warn, block, or review verdicts.
• It includes shell deobfuscation, SafeFix suggestions, and RiskChain detection for multi-step attacks.
• An LLM-as-Judge evaluates ambiguous inputs with low-millisecond latency.
• On adversarial scenarios, AgentTrust achieved 96.7% verdict accuracy, including 93% on obfuscated shell payloads.

Why it matters

For startups integrating AI agents, runtime safety is paramount. AgentTrust offers a practical, low-latency solution to prevent catastrophic failures from unsafe tool use, protecting both the startup's operations and its users.

💊 AI Business

Curated Data Outperforms LLMs in Drug Discovery

The Rundown

General-purpose LLMs with web search are increasingly used for scouting pharmaceutical pipelines. However, their effectiveness is limited, especially for niche targets in preclinical or Asian-developed assets. A new AI platform, Gosset, challenges this. Gosset uses a chat interface backed by curated target-, modality-, and indication-level drug-asset annotations. Researchers benchmarked Gosset against four frontier LLMs—Claude Opus 4.7, GPT 5.5, Gemini 3.1 Pro, and Perplexity sonar-pro—on ten niche oncology/immunology targets. All systems received identical natural-language queries and JSON output schemas. Gosset returned 3.2 times more verified drugs per query than the best frontier system. Crucially, Gosset achieved perfect precision and 100% recall against the union of verified drugs found by all systems. This suggests that a curated index, exposed as a tool for LLMs, can significantly close recall gaps compared to generic web search.

The details

• Gosset, an AI platform with curated drug-asset annotations, outperformed frontier LLMs in pharma discovery.
• On 10 niche oncology/immunology targets, Gosset returned 3.2x more verified drugs than the best LLM.
• Gosset achieved perfect precision and 100% recall against the combined findings of all tested systems.
• A curated index, callable as a tool by LLMs, can bridge recall gaps from generic web search.

Why it matters

Startups in biotech and pharma can leverage curated data platforms like Gosset to accelerate drug discovery. This approach offers a significant competitive advantage over relying solely on general-purpose LLMs for competitive intelligence and asset scouting.

🎤 NLP

VocalParse: Singing Voice Transcription

The Rundown

High-quality singing annotations are essential for Singing Voice Synthesis (SVS) systems, but manual labeling is labor-intensive and requires musical expertise. Automatic annotation is crucial for scalability. Current systems often use complex multi-stage pipelines, struggle with text-note alignment, and generalize poorly to out-of-distribution data. VocalParse addresses these issues. It's a unified singing voice transcription (SVT) model built on a Large Audio Language Model (LALM). VocalParse introduces an interleaved prompting formulation that jointly models lyrics, melody, and word-note correspondence, directly generating a structured musical score. It also employs a Chain-of-Thought (CoT) style prompting strategy, decoding lyrics first as a semantic scaffold. This mitigates context disruption while preserving interleaved generation benefits. Experiments show VocalParse achieves current best SVT performance on multiple singing datasets.

The details

• VocalParse is a unified singing voice transcription model based on a Large Audio Language Model (LALM).
• It jointly models lyrics, melody, and word-note correspondence using interleaved prompting.
• A Chain-of-Thought (CoT) strategy decodes lyrics first to mitigate context disruption.
• VocalParse achieves current best performance on multiple singing voice transcription datasets.

Why it matters

For startups in music tech, AI-powered music generation, or karaoke applications, VocalParse offers a significant leap in transcription accuracy. This tool can streamline content creation and enhance user experiences in the burgeoning AI music market.