Block unveils Dessa, a speech-only AI that talks and listens at once

Block unveils Dessa, a speech-only AI that talks and listens at once

Block's research team has introduced Dessa, a full-duplex speech-to-speech model trained from scratch on recorded human conversation, without a text-training stage or a pretrained language model, according to the company. It can speak while listening, including interruptions and short backchannel responses. Block says it runs faster than real time on a MacBook Pro with an M4 Max chip or better. Saturday's reporting and demonstrations attracted attention for natural timing and expressive speech.

Why it matters. Speech-native AI could preserve timing, tone and emotion that transcription-first systems may discard, and could eventually make local voice agents more practical. For now, Dessa is a research preview. Block has not published downloadable weights, an API, standardized benchmarks or an independent evaluation. The impressive clips are selected demonstrations, not proof of general reliability.

Independent Kimi K3.1 Pro testing puts the new model near the frontier

Fresh October 24 analysis of Artificial Analysis measurements places Moonshot's Kimi K3.1 Pro at roughly 55 on the Intelligence Index, fourth among the evaluated models at their highest reasoning settings. A reported Terminal-Bench 4.0 score of 67% compares with 13% for Kimi K3, indicating a large gain in this agentic coding test. The 2.8-trillion-total-parameter mixture-of-experts model is currently available through Moonshot's services, with downloadable weights promised by November 4.

Why it matters. This is a substantive independent follow-up to yesterday's launch, rather than another repeat of Moonshot's marketing claims. Composite benchmarks are sensitive to methodology and reasoning settings, and performance on a benchmark is not a guarantee on a real codebase. The weights and final license are still unavailable, so claims about its future open-weight standing remain conditional.

Italian cyber agency warns that a third PHP flaw remains unpatched

In an October 24 advisory, Italy's CERT-AGID clarified that CVE-2026-115560 is separate from the two critical PHP vulnerabilities patched on October 21. This newly disclosed remote-code-execution path requires an unusual combination of PHP session settings and application behavior. The PHP team does not plan an emergency release, and the fix is expected in the next routine update. CERT-EU also added the issue to its October 23 advisory.

Why it matters. Administrators should distinguish the widely exposed PHP2Shell flaws, for which updates are available, from this narrower issue. Installing the October 21 patches alone does not address CVE-2026-115560. For affected configurations, the published temporary mitigation is to disable PHP's session upload progress feature and restart the relevant workers. No widespread exploitation of this new issue has been established.

New scrutiny of OpenAI's monitoring gaps shows why outside audits matter

Weekend follow-up reporting examines a high-access review published October 22 by METR and Redwood Research. Their researchers inspected an earlier version of OpenAI's internal agent-monitoring infrastructure and identified ways for agent actions to escape oversight, including a method that could hide hundreds of recent actions. Coverage estimates for eligible internal agentic Responses API traffic ranged from roughly 75% to 94%, with substantial uncertainty. OpenAI says production-relevant vulnerabilities have been patched.

Why it matters. Monitoring is useful only if it reliably sees the actions it is supposed to evaluate. The audit demonstrates the value of external researchers having technical access to internal systems. The reviewed configuration dates to July, not the current production setup, and METR explicitly says it has not verified OpenAI's later remediation claims.

A new AI shutdown tally tracks what happens when products lose their edge

TechCrunch published a new October 24 survey of abandoned and troubled AI products, including workflow-automation startup Relay, which closed in August, and Character.ai, which filed for bankruptcy earlier in October. The feature also looks at discontinued or absorbed products from larger labs, showing that access to capital and technical novelty have not guaranteed lasting demand.

Why it matters. As general-purpose assistants absorb specialist features, a standalone AI product needs repeat use, reliable economics and a defensible reason to exist. This is a retrospective analysis published yesterday, not a claim that all of these shutdowns happened in the last 24 hours. The cases differ substantially, and they should not be treated as evidence that all AI startups face the same failure rate.

Watch for independent Dessa evaluations and actual local downloads, the promised November 4 release of Kimi's weights and license, the next PHP maintenance update, and a third-party recheck of OpenAI's revised agent-monitoring safeguards. The next week will show which claims turn into reproducible evidence.