Good evening. Here is what matters in AI today, and how to put it to work.
AI agents are becoming load-bearing infrastructure: Claude now reads your bank account, Perplexity fields workplace teammates, and critical supply-chain CVEs follow right behind.
~4 min read · last 12 hours
In today's issue
01
Perplexity launches always-on 'Teammates' for workplace AI agents
02
Claude can now connect directly to your bank account via Plaid
03
Airbnb plans an AI travel agent for the first half of 2027
04
Amazon Bedrock AgentCore adds single-call Web Answer with built-in citations
05
Critical: kaniko can leak Azure access tokens to attacker-controlled registries
Main story
Perplexity launches always-on 'Teammates' for workplace AI agents
Perplexity's new Teammates agents take on persistent roles inside a workspace, drawing on shared projects, skills, and memory to compete directly with OpenAI and Anthropic in the enterprise.
Why it matters: Persistent, role-aware agents with shared memory are a meaningful architectural step beyond single-session chatbots, and this raises the bar for any enterprise AI roadmap still built around one-off prompts.
What to watch next: Watch whether Perplexity's Teammates gain traction in organizations that already use Perplexity for search, since that existing footprint is the fastest path to enterprise stickiness against OpenAI and Anthropic.
From Perplexity's workplace teammates to Claude managing your bank account, Airbnb booking your next trip, and Amazon Bedrock handling web answers in a single call, we are watching AI agents shift from experimental features into load-bearing parts of real products and platforms.
CoreWeave Registry: The System of Record for Enterprise AI Assets
CoreWeave
Dev Community Live: NVIDIA Developer Live Agent Gauntlet
NVIDIA Developer
The Signal
The cluster of agent launches today is not a coincidence: Perplexity, Anthropic, Airbnb, and AWS all moved on the same week, signaling that the race to own the "always-on agent" layer is now fully open. For engineering and product leaders, the practical question is no longer whether to build agentic workflows but how to govern them once they hold live credentials, persistent memory, and the authority to act. At the same time, the kaniko and banks vulnerabilities show that AI-adjacent infrastructure is attracting the same attacker attention as the models themselves, so security review of the full agent stack, not just the model layer, is now non-negotiable.
All the best, the KYFEX team
“Perplexity is taking on OpenAI and Anthropic with its own work-focused AI agents.”
The Verge
Quick hits
AI agents move from demo to daily infrastructure
Claude can now connect directly to your bank account via Plaid
Anthropic's new Personal Finance feature lets Claude read transaction data through Plaid so users can ask natural-language questions about their spending.
Why it matters: Giving a general-purpose AI agent read access to financial data is a trust and liability threshold most enterprises have not yet crossed, and how Anthropic handles this will set expectations across the industry.
Airbnb plans an AI travel agent for the first half of 2027
Airbnb's upcoming agent will handle trip planning, option comparison, booking management, and in-stay support, positioning it as an end-to-end travel concierge.
Why it matters: A major consumer platform committing to a full-lifecycle agent by mid-2027 signals that agentic UX is now a competitive necessity, not a differentiator.
Amazon Bedrock AgentCore adds single-call Web Answer with built-in citations
AWS's new Web Answer tool on Bedrock AgentCore returns a synthesized, source-cited answer to a natural-language question in one call, eliminating the need to build a search-retrieve-synthesize-cite pipeline manually.
Why it matters: Collapsing a four-step RAG pipeline into a managed single call cuts both build time and failure surface, making it a practical default to evaluate before rolling your own retrieval stack.
OpenAI shipping a dedicated cybersecurity model and a critical kaniko vulnerability leaking Azure tokens to attacker-controlled registries land on the same day, a reminder that AI capability advances and AI-adjacent infrastructure risks are moving in lockstep.
Critical: kaniko can leak Azure access tokens to attacker-controlled registries
Versions of kaniko before v1.25.20 will send a Microsoft Entra ID token to any lookalike registry host referenced in a Dockerfile, giving an attacker full credential access.
Why it matters: Any CI/CD pipeline using kaniko to build container images in Azure environments should treat this as an urgent patch, since a single malicious or misconfigured Dockerfile reference is enough to expose cloud credentials.
Prototype and deploy GenAI apps using an MVP workflow, prompt engineering, and RAG.
DeepLearning.AI · Free · 1 hour
Put it to work
Try this today
Audit an agentic Python workflow for RCE risk via dynamic imports
You are a security-focused code reviewer. I will paste a Python file that uses an LLM completion library. Your job is: 1. Identify every place where a string value derived from LLM output, user input, or a rendered template is passed to importlib.import_module(), getattr(), eval(), or exec(). 2. For each finding, rate the exploitability (high / medium / low) and explain why. 3. Suggest the minimal code change that eliminates the risk without breaking the intended functionality.
Here is the file: [PASTE YOUR CODE HERE]
Why it helps: The banks RCE disclosure today shows exactly this pattern in a widely used LLM templating library, making this audit worth running on any agentic codebase before your next deploy.
Before you ship it
The risk
Connecting a general-purpose AI agent to live financial accounts via OAuth bridges like Plaid creates a broad, hard-to-audit permission scope: a prompt injection in retrieved transaction data could instruct the agent to exfiltrate or misrepresent account details.
Do this
Scope the Plaid integration to the minimum required accounts and transaction history window, enforce read-only token permissions at the API level, and add a human-review step before the agent surfaces any actionable financial recommendation.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
