In Today’s edition:

1. Anthropic Reveals Claude Hacked Three Companies During Security Tests

2. PMNA TUTORIAL: How to Build an AI Load Balancer for Your Automation Backlog

3. Atlassian Tightens AI Spending as Tokenmaxxing Gains Momentum

4. TOOL REVIEW: Murmell Helps Teams Run Multiple AI Coding Agents on One Repository Without Overwrites

5. Exciting Career Opportunities for Product and Project Management Professionals

Reading time: 5 minutes

HEADLINE NEWS

Anthropic Reveals Claude Hacked Three Companies During Security Tests

Anthropic has disclosed that several versions of its Claude AI model unintentionally gained unauthorized access to three real-world organizations while undergoing internal cybersecurity evaluations. The company described the incidents as an "operational failure", stressing that they resulted from testing environment mistakes rather than deliberate malicious behavior or failures in the models' alignment.

• Claude compromised three separate organizations during simulated cybersecurity "capture-the-flag" evaluations.

• The affected models included Claude Opus 4.7, Claude Mythos 5, and an internal research model.

• Rather than exploiting sophisticated zero-day vulnerabilities, the AI leveraged basic security weaknesses such as weak passwords and exposed endpoints.

• The incidents occurred because the testing environment was mistakenly connected to the public internet without the normal safeguards Anthropic uses for deployed models.

• Anthropic has notified the affected organizations, launched a review of its testing procedures, and says the events highlight the need for stronger governance around increasingly capable AI agents.

The incidents underscore how rapidly AI systems are advancing in cybersecurity tasks—and how important robust testing controls have become. While Anthropic says the breaches were caused by operational mistakes rather than intentional AI misconduct, the episode is likely to intensify industry and regulatory scrutiny over how powerful AI agents are evaluated before deployment. Read More 

ONE MONTH FREE OF MAKE.COM!

If you’ve been looking for a smarter way to automate your workflows, now’s the perfect time to try Make. It’s one of the most powerful tools for connecting apps, streamlining tasks, and building systems that run without you.

We have brought you a fine solution. This gives you full access to advanced features to enable you to automate lead gen, content distribution, or internal operations.

For a limited time, you can get 1 month of the Pro plan completely free using my link below:

PMNA TUTORIAL: How to Build an AI Load Balancer for Your Automation Backlog

This PMNA tutorial builds a recommendation engine that reads two monday.com boards (an intake board of SME-submitted automation requests and a delivery team board of people, skills, and capacity) and returns a ranked backlog, a staffing proposal, and drafted replies for rejected submissions. It exists because intake piles get scored inconsistently and staffed by whoever argued hardest. The hard constraint throughout: Claude recommends, a human assigns.

Key Steps and Ideas:

• Build both boards, including the decision-points field that exposes fake automation candidates.

• Fix the five scoring criteria: volume, rules-based vs. judgment, exception rate, systems touched, time saved.

• Prompt Claude to score every submission 1–5 per criterion, total, rank, and flag bad candidates with a one-line reason.

• Follow up with build effort (S/M/L) and a quick-win flag, then split the list into quick wins and big bets.

• Prompt for staffing on the top 3, matching skill tags against active projects and capacity status — flag the conflict rather than assigning to someone already Full.

• Prompt for under-150-word draft emails to SMEs whose ideas were declined, referencing their actual process and a smaller automatable slice.

• Respect the failure modes: thin intake data, stale capacity data, treating the score as the decision, and wiring it to auto-assign.

What You Should Consider:

The whole thing runs on the same three-line prompt structure (context, task, constraints) pointed at two boards instead of one. The output isn't an answer so much as a ranked argument you can read and overrule, which is what keeps it safe to run against a real team. Its ceiling is entirely set by your board hygiene.

FULL GUIDE HERE

Atlassian Tightens AI Spending as Tokenmaxxing Gains Momentum

Atlassian is introducing stricter controls over how employees use AI by assigning monthly AI spending budgets, or "AI wallets," to staff. The move comes as many technology companies grapple with soaring AI costs, while others are encouraging employees to maximize AI usage… a trend known as "tokenmaxxing."

• Atlassian has introduced monthly AI wallets ranging from $500 to $2,000 for employees in its R&D teams.

• Employees receive alerts as they approach their AI spending limit, and access pauses until additional funds are approved.

• The company says the wallets are designed to encourage responsible AI experimentation rather than unrestricted usage.

• Other tech firms have embraced "tokenmaxxing," with some reportedly tracking AI usage through internal leaderboards, though companies like Uber and Amazon have also faced unexpectedly high AI costs.

• Industry analysts believe AI spending caps help reduce waste, especially as autonomous AI agents dramatically increase token consumption and operational costs.

The story highlights a growing shift in enterprise AI adoption, from encouraging maximum experimentation to managing AI as a measurable business expense. As AI agents become more powerful and expensive to operate, companies are increasingly balancing innovation with tighter cost controls and governance. Read More 

TOOL REVIEW: Murmell Helps Teams Run Multiple AI Coding Agents on One Repository Without Overwrites

Murmell puts your entire team and every coding agent you run into a single live cloud workspace, building on the same repository at the same time. Agents reserve the files they're about to edit before they write, so multiple people and multiple agents can work in parallel without overwriting a line of each other's work. Close your laptop and the agents keep going — the work is still moving when you return.

• Claim-before-write reservations — agents lock the files they're about to edit, turning parallel work into a coordinated build instead of a merge cleanup.

• One repository, one room — every agent works in the same directory and sees teammates' changes as they land, not at merge time.

• Bring your own agents — Claude Code, Codex, Kimi and OpenCode run side by side today on the accounts you already pay for, with more on the way.

• Share a link, share the room — anyone can join the canvas and watch the terminals run live, with names and cursors on screen.

• Everything lands in git — each project pushes continuously to its own private repository, so the machine is disposable, and the code never is.

Ask a team what their agents are working on right now and the honest answer is usually "let me check." Murmell replaces that with a live view of the work itself: who is active, which files are locked, and what every agent is producing at this moment. There are no standups to chase, no screen shares to schedule, and no waiting on a pull request to discover a decision that was made three hours earlier. Send the link to a stakeholder, and they watch progress happen rather than hear about it next week. For anyone accountable for delivery, that is the difference between reporting on a project and actually seeing it move.

TOOL LINK

Exciting Career Opportunities for Product and Project Management Professionals

• Implementation Specialist @ Johnson Controls

  📍 Raleigh, NC.

• Lead Product Manager - Tamarac @ Envestnet

  📍 Raleigh, NC.

• Senior Lead Product Manager - Payments Transformation, Value Added Services @ Wells Fargo

  📍 Charlotte, NC.

• Project Manager @ ABM Industries

  📍 Garner, NC.

• Senior Program Manager - DERMS @ Itron

  📍 North Carolina.

THAT’S A WRAP