Daily AI: Block’s voice preview, Fortinet warning and new building tools

Block shows Dessa, a voice AI that keeps listening

Block, the company behind Square and Cash App, announced Dessa on 23 October. It is a research preview that can listen while speaking, according to the company.

• Block says it trained the model on conversation audio rather than text or a pretrained model.
• Its recordings demonstrate interruptions, timing and expression. They are company examples, not independent performance tests.
• Block says it runs locally on a MacBook Pro with an M4 Max chip or better. That is not a claim about every Mac.
• Early access is for a small group of safety researchers. No public release date, API, downloadable weights or price was established from the announcement.

For a business, this is a development to watch. A convincing recording does not establish that you can connect it to customer calls today.

CERT-In issues a new Fortinet warning

India’s CERT-In issued CIAD-2026-0054 on 24 October. It urges action on Fortinet security flaws found through AI-assisted testing.

• The affected product families include FortiOS, FortiManager, FortiAnalyzer, FortiMail and FortiAuthenticator. Exposure depends on the installed product and version.
• The vendor announcement is dated 21 October, with coordinated updates planned for 22 October. The new event here is the 24 October CERT-In advisory.
• The warning does not establish that your business has been attacked. Detailed vulnerability identifiers are due on 10 November, according to the sources.

1. Identify any affected products, especially devices exposed to the internet.
2. Have the responsible IT team back up configurations and plan the appropriate update.
3. If patching cannot happen immediately, follow the vendor’s temporary access restrictions; these do not replace patching.

Ask the person responsible for your network to check the official version table. Do not guess an upgrade from a social-media headline.

NVIDIA releases a smaller model for device-based assistants

NVIDIA published the Nemotron 3.5 Bolt E4B 4GB release on 23 October. This is the actual release of a model previously announced for that date.

• It targets conversations and tool use on supported NVIDIA systems, including GeForce RTX and Jetson devices.
• The roughly 4GB description refers to target workloads and configurations. The published measurements vary; it is not a universal total-memory requirement.
• A voice application still needs speech software around the language model. The company’s benchmark gains are not independent results.

Developers can assess the released model and deployment notes. Business owners should ask for a test on their real task and intended hardware.

Ollama’s latest release supports spoken replies on macOS

Ollama’s official v0.40.11 release is dated 24 October. Its notes describe generating speech on macOS and streaming audio to applications.

• With a supported speech model, a reply can play as its audio is produced. This is not a promise that every installed model can speak.
• The release also describes d3 decision models: models that sort information into choices, yes-or-no answers or scores.
• These are features described in the current release, not a claim that they first appeared on 24 October. Compatible models and hardware still matter.

For example, a developer could test whether a short support update is spoken clearly. We have not run or benchmarked this release.

TanStack AI reaches 1.0 with a stable core

TanStack announced AI 1.0 on 23 October. It is a toolkit developers can use to add AI chat, tools and media features to an application.

• The core interfaces are stable. The new agent harness, which coordinates ongoing agent work, remains experimental.
• Supported adapters connect different model providers. This offers choices but does not make every provider’s features identical.
• Existing users should follow the upgrade guide, including changes to stored conversations. Stable packages move together; experimental harness packages remain on 0.x.

It is a building tool rather than a finished customer-support product. Allow time for integration and testing.

Cisco adds tests for multi-step security agents

Cisco introduced Agentic FAITH on 23 October. It extends an evaluation framework to agents that choose tools and investigate software over several steps.

• Its first integrated benchmark, VulnLoc, tests whether an agent finds vulnerable files.
• It also checks whether the agent avoids reporting the same problem in patched code.
• Saved tool calls and results help teams examine how an answer was reached.

For buyers, the useful question is whether a security agent can show its evidence and false alarms, not just a confident final answer. This framework does not certify a system as secure.

Palo Alto security offerings reach OpenAI Marketplace

Palo Alto Networks announced the marketplace availability of Unit 42 Continuous Frontier AI Defense and Prisma AIRS on 23 October.

• These are enterprise security offerings, covering continuous assessment and protections for AI systems.
• Eligible OpenAI enterprise spending commitments may apply under the commercial terms. This is not a free security switch in consumer ChatGPT.

A new buying route is not independent evidence that a product will protect every workflow. Check scope and cost for your environment.

Virginia orders an AI-cyberattack preparedness plan

The US state of Virginia announced Executive Directive 6 on 23 October, directing a statewide cybersecurity plan and coordination effort.

• The directive addresses AI-related cyber risks and calls for regular expert briefings.
• It refers to earlier agent activity, including fake information requests. Those earlier events are not newly dated attacks.
• The governor’s office said it understood that no sensitive information was accessed. That is the office’s assessment.

This is regional policy news, not a new requirement for companies in India or the GCC.