Agents move into Slack as prompt injection stays unsolved

Agent security spans inputs and installed capabilities

Prompt injection remains a problem for agents that ingest untrusted content. Willison frames it as role confusion: models cannot reliably distinguish instructions from data. Latent Space's piece on Gray Swan covers indirect prompt injection techniques and defenses used in practice. These accounts suggest current mitigations reduce exposure without resolving the underlying problem.

Risk also arrives through installed capabilities. Unit42 at Palo Alto Networks describes how OpenClaw's marketplace of pre-built agent skills creates a supply-chain attack surface similar to npm or PyPI, involving capabilities that can take real actions. NVIDIA outlines a toolkit pairing open models, tools and skills with a secure runtime for specialized enterprise agents. Separately, the Guardian reports on research tied to Anthropic's Claude and the Fable AI model suggesting systems capable of devastating attacks on governments and businesses could arrive within months. That is a forecast, rather than an established capability.

Agents enter conversations and application stacks

Claude Tag lets teams @-mention Claude in Slack to assign tasks and connect it to internal tools and codebases, keeping work inside the conversation. That may lower adoption friction for teams already coordinating in Slack. The supported workflow centers on explicit task assignment; it does not establish broader access to channel traffic.

For teams building their own agents, IBM Research's CUGA offers two dozen working agentic app examples on a lightweight harness, providing concrete reference implementations for multi-step tasks. n8n's 2026 report on agent development tools surveys choices between low-code orchestration, code-first frameworks and hybrid approaches. At the infrastructure level, NVIDIA and AWS describe a collaboration on bringing AI workloads to production at scale, covering cloud infrastructure and tooling.

Document parsing gets a new baseline

Mistral claims state-of-the-art document intelligence performance for OCR 4, targeting layout understanding, table extraction and mixed text and image documents. Those capabilities address a concrete bottleneck in retrieval and data-extraction pipelines: turning source documents into usable structured input. A stronger OCR baseline could reduce reliance on brittle parsing heuristics or proprietary document APIs, though the performance claim remains Mistral's.

Longer video, technical detail and planned voice

ByteDance's Seedance 2.5 can generate clips up to 30 seconds long from a single text prompt, extending the length available for ad, social and entertainment workflows. For engineers interested in how generative models are built, Krea published a technical report detailing Krea 2's architecture, training methodology and design decisions. The report offers implementation detail alongside the day's product announcements.

Voice interfaces have a prospective change of their own: OpenAI is reportedly preparing a bidirectional voice mode for ChatGPT that would support more natural, interruptible conversation instead of turn-based exchanges. For teams building voice-driven agents, that makes conversational turn handling a relevant comparison point, while the rollout remains a reported plan.