Skip to content
DailyPrompt

your daily dose of claude

01 product

Claude Managed Agents Learn to Dream, Spawn Specialist Sub-Agents and Set Their Own Success Criteria

At Code w/ Claude 2026, Anthropic shipped three big upgrades to Claude Managed Agents. A research-preview 'dreaming' mode reviews up to 100 past sessions to tidy memory and surface recurring patterns, multi-agent orchestration lets a lead agent delegate work to as many as 20 specialists running in parallel, and outcomes lets developers define explicit success criteria so Claude iterates without constant hand-holding.

9to5Mac →
product

Claude Lands Inside Excel, PowerPoint and Word as Microsoft 365 Add-Ins Hit General Availability

Claude is now generally available as add-ins for Excel, PowerPoint and Word, with an Outlook beta rounding out the suite. Conversation context follows you between the Microsoft 365 apps, the add-ins are available on all paid plans, and admins can deploy them from the Microsoft admin center or route traffic through Amazon Bedrock, Google Vertex AI or Microsoft Foundry.

Anthropic
C
Claude
@claudeai

Claude for Excel, PowerPoint, and Word are now generally available, and Claude for Outlook is in public beta. As Claude moves between your Microsoft apps, it carries the full context of your conversation.

12h View on X →
product

$ cat story_03.md

Claude Code Now Reviews Every Pull Request With a Team of Bug-Hunting Agents

Anthropic has launched Code Review for Claude Code, the same multi-agent pipeline it already runs on nearly every pull request internally. When a PR opens, Claude dispatches a fleet of agents that hunt bugs in parallel, verify each finding to weed out false positives, then post a single ranked overview comment with in-line annotations on the issues worth fixing.

$ open anthropic →
research No. 04

Anthropic Teaches Claude to Translate Its Own Thoughts Into Plain English

New interpretability research introduces Natural Language Autoencoders, a pair of trained modules that turn Claude's internal activations into readable text and back again. In alignment audits, an investigator equipped with NLAs uncovered a target model's hidden motivation between 12 and 15 percent of the time without ever seeing the training data, and Anthropic has released training code and trained NLAs for popular open models.

Source: Anthropic →
A
Anthropic
@AnthropicAI

We're donating Petri, our open-source alignment tool, to @meridianlabs_ai, so its development can continue independently. Working with Meridian Labs, we've also released a major update that improves the adaptability, realism, and depth of Petri's tests.

9h View on X →
05
research

Anthropic Hands Petri Alignment Tool to Non-Profit Meridian Labs

Anthropic has transferred Petri, its open-source automated alignment auditing framework, to Meridian Labs, a 501(c)(3) non-profit, so the tool can continue evolving outside the company. The handover ships with a fresh Petri update, co-developed with Meridian, that improves the realism, adaptability and depth of its safety tests.

The Rift →
announcement

Anthropic Opens Its Security Bug Bounty Programme to the Public on HackerOne

After running privately within a closed pool of security researchers, Anthropic's security bug bounty programme is now public on HackerOne and accepting reports from anyone. The move broadens the talent pool hunting for vulnerabilities across Claude and the rest of Anthropic's product stack, with rewards scaled to severity.

Read on HackerOne →
Breaking

Live Notes From Code w/ Claude 2026: Routines, Remote Agents and a 17x Surge in API Volume

Simon Willison's live blog captures the firehose of announcements from Anthropic's San Francisco developer conference: doubled Claude Code rate limits, automatic fixes on CI pull requests, security reviews, 'routines' for asynchronous PR-ready automations and remote agents you can steer from a phone. Anthropic also said API usage on the platform is up 17 times year on year.

Simon Willison's Weblog
A
Anthropic
@AnthropicAI

Our security bug bounty program is now public on HackerOne. We've run the program privately within the security research community, and their findings have strengthened our products. Now anyone can report vulnerabilities and get rewarded.

11h View on X →
A
Anthropic
@AnthropicAI

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers, called activations, encode Claude's thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.

13h View on X →
08 announcement

Fake Claude AI Site Pushes 'Beagle' Backdoor to Windows Users via Sponsored Search Ads

Researchers have flagged a counterfeit Claude AI site distributing a 505MB MSI installer dressed up as 'Claude-Pro Relay'. Behind the install chain sits Beagle, a previously undocumented Windows backdoor that sideloads through a legitimately signed G Data DLL and beacons home over encrypted TCP and UDP. Attackers are pushing the fake pages to the top of search results with paid Google Ads.

BleepingComputer →
A
Anthropic
@AnthropicAI

We're sharing the research agenda of The Anthropic Institute, or TAI. TAI will focus on four areas: 1) Economic diffusion 2) Threats and resilience 3) AI systems in the wild 4) AI-driven R&D

16h View on X →