AI Pulse

Zendesk Kills Its Old Bots in Six Days. There Is a Patched Reasoning Token Flaw in Every Major AI API. The August 25 Pulse.

By Felix Maru · August 25, 2026 · 6 min read

Six days. On August 31, Zendesk stops active development on its legacy bot builder and Answer Bot framework. On December 10, the plug comes out entirely, and any support workflow still running on the old system stops working. That is the urgent item this week. But two other stories from the past few days also deserve your attention: a patched but genuinely instructive security flaw that hit all three major AI API providers at once, and Anthropic's first ever operating profit, which tells you something real about how the AI market has shifted.

The Six-Day Warning: Zendesk's Legacy AI Is Entering Maintenance Mode

Zendesk is retiring its older AI layer in two stages. August 31 is when active development stops for AI agents - Essential and the legacy bot builder (the old intents system and classic Answer Bot). After August 31, only critical bug fixes will be applied to that infrastructure. December 10 is the hard deadline: legacy bots stop functioning entirely on that date.

What is being retired: the original bot builder interface, the intents-based Answer Bot, and the AI agents - Essential tier. If you built your Zendesk bot using the old "Answer Bot" configuration screen or the bot builder that asks you to map intents manually, you are on the affected platform.

What replaces it: the new AI agent experience, which includes agentic reasoning, multi-step procedures, and external API integrations. The practically important change here is that the features previously locked behind the "AI agents - Advanced" add-on are now included across all Suite and Support plans. If you have been waiting to upgrade because the advanced features cost extra, the migration actually unlocks more capability at what you are already paying.

How to check if you are affected: in your Zendesk admin, navigate to the AI agents section. If your interface shows the old bot builder layout or an "Essential" label, you are on legacy infrastructure. Zendesk published a migration guide at their support site. Third-party migration guides from Help Desk Migration and BDQ Cloud cover the mechanics in more detail.

My read: the August 31 date matters more than December 10 for one practical reason. Once development stops, any bugs or edge cases you find in your existing bot will not get fixed. If your team runs through a high-volume period before December (product launches, seasonal spikes, any campaign), you will be running on an unmaintained system. Start the migration now, not in November.

The Reasoning Token Flaw You Should Understand Even Though It Has Been Patched

Earlier this month, researchers at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems published a paper with an uncomfortable finding. OpenAI, Anthropic, and Google all used a single global encryption key to protect the hidden "reasoning tokens" that their thinking models generate before producing a visible response.

Using that shared key, the researchers decoded 315,320 hidden reasoning blocks. In those blocks, pulled from publicly available API logs and shared transcripts, they recovered 182 real credentials: API keys, session tokens, and passwords that had ended up inside reasoning traces without anyone realizing they were there. The main extraction attack is no longer reproducible. All three providers updated their key management after the paper published.

The lesson is worth holding even after the patch. When you call Claude, GPT, or Gemini via API with extended thinking enabled, the response may include a reasoning block alongside the visible text. If you log raw API responses for debugging (and most teams building support automations do), those logs contain reasoning trace data. If those logs end up in a shared repository, a ticket system, a logging dashboard others can access, or a public GitHub repo, you are potentially exposing content that is denser and more sensitive than the visible reply.

The fix: strip reasoning blocks and opaque reasoning fields before storing or sharing API transcripts. Do not commit raw API call logs even when the visible output looks harmless.

This applies to any team using extended thinking in Claude, o-series models in OpenAI's API, or Gemini's thinking mode. The flaw is patched at the provider level. The hygiene practice is yours to implement. Sources: Cybersecurity News, ExplainX, Welcome.AI.

Anthropic's First Profitable Quarter (And Why It Will Not Last, and Why That Is Fine)

Anthropic's Q2 2026 financials show roughly $10.9 to $11.5 billion in revenue and approximately $559 million in operating profit. That is the company's first profitable quarter, roughly two years ahead of its own projections. Revenue grew roughly 130% compared to Q2 2025.

Before you read this as a clean profitability milestone: Anthropic itself told investors the margin will not hold. Significant infrastructure investment is planned. Some analysts have noted that temporary ramp-up discounts on compute contracts inflated Q2's margin, and those discounts will not be permanent. AI Weekly and Forbes both covered this with useful context.

My honest read: the quarter's one-time profitability is less interesting than the revenue scale. A company generating north of $10 billion per quarter at 130% year-over-year growth has moved past the "interesting experiment" phase into something that looks like infrastructure. It does not make bets on pilots and hobbyist projects at that revenue level. It has real enterprise customers signing real contracts.

For anyone building support automation on Claude: this is a stability signal. The company's long-term financial position is still uncertain (large infrastructure investments ahead, compute costs are real), but the commercial foundation is now visibly solid. The platform is not going anywhere.

AI Agents in Customer Service Just Crossed 60%: What the Adoption Data Actually Shows

A Futurumgroup report published this month puts AI agent deployment in customer service at roughly 66% of enterprises surveyed, up from roughly 39% in 2025. Roughly 70% of those that have deployed report measurable value within 60 days. The top improved key performance indicator, ahead of cost savings and ticket deflection volume, is customer satisfaction.

That last point is the one worth sitting with. CSAT as the leading benefit, not cost per ticket or headcount reduction, suggests that the deployments working best are not the ones that removed human agents from the channel. They are the ones where AI took the repetitive layer (triage, lookups, draft replies, the simplest tier-1 resolutions) so that human agents have more capacity for cases that require judgment, context, and trust.

The teams I have seen struggle with AI agents in support are typically the ones who started with the goal of reducing headcount. The ones seeing CSAT gains started with the goal of making their existing team faster and less burned out. Customers can tell the difference between an AI that handles the routine so a human can take the hard thing, and an AI that exists so no human has to be available at all. The data is now large enough to make that pattern visible.

If your team is still in a cautious pilot, you are not early anymore. The majority has moved. But "deployed" is not the same as "deployed well." The gap between a thoughtfully designed human-plus-AI support model and a poorly designed one is larger now than it was 18 months ago, because customer expectations have shifted upward. Getting the design right matters more, not less.

Sources

Working through any of the moves above, or thinking about how to structure a human-plus-AI support model that actually improves CSAT? Drop me a note and I'm happy to compare notes on what I've seen work.

Share 𝕏 in ✉

Comments