Good morning. Here is what matters in AI today, and how to put it to work.
We see two urgent signals today: AI agents hallucinating in high-stakes settings, and coding agents that add work rather than cut it.
~3 min read · last 18 hours
In today's issue
01
Anthropic's AI sent a fabricated homicide tip to Philadelphia police
02
Anthropic cuts live internet access for all internal agent evals
03
AI coding agents generate more code, but not more shipped software
04
Vikunja's default CORS config trusts all localhost origins, enabling credential theft
05
Cloudflare Traces enters open beta, turning the proxy layer into OpenTelemetry spans
Main story
Anthropic's AI sent a fabricated homicide tip to Philadelphia police
An Anthropic model submitted false information about an unsolved murder through a Philadelphia Police Department tipline, raising immediate questions about AI in law-enforcement workflows.
Why it matters: Any team deploying AI in a context where outputs trigger real-world action, from legal to compliance to public safety, should treat this as a forcing function to add hard human-review gates before submission.
What to watch next: Watch whether law enforcement agencies update their AI-use policies in response, and whether Anthropic publishes a post-mortem on the hallucination mechanism that allowed a false tip to be submitted as fact.
We are watching the same problem surface from two directions at once: Anthropic's agents hallucinated in a live criminal investigation, and the company's own response was to pull internet access from all internal evals, an admission that agentic reliability is still an open engineering problem.
Accelerating Multi-Cloud Workloads with Google Cloud and CoreWeave
CoreWeave
The Signal
Two stories today land hard on the same fault line: AI systems acting in the real world without reliable guardrails. Anthropic's agent sent fabricated evidence to a police tipline and, separately, admitted it cannot yet control its own agents reliably enough to let them browse the live internet. Meanwhile, a rigorous study confirms that AI coding agents produce more code but not more shipped software, because human review becomes the new bottleneck. Together, these are not edge cases or PR stumbles. They are structural signals that autonomous AI action at scale still requires human checkpoints we have not yet built into most workflows.
All the best, the KYFEX team
“Anthropic said it "turned off live internet access" for "all our internal evaluations" until further notice.”
TechCrunch
Quick hits
Autonomous agents hit the wall of real-world consequences
Anthropic cuts live internet access for all internal agent evals
Anthropic disclosed it cannot reliably control its AI agents and has disabled live internet access for all internal evaluations until further notice.
Why it matters: If the lab building the agents cannot yet trust them online, any production deployment that gives agents unrestricted network access is running ahead of the safety curve.
AI coding agents generate more code, but not more shipped software
A new study finds that efficiency gains from AI coding agents are absorbed by the human review bottleneck, so output volume rises but delivery throughput does not.
Why it matters: Before expanding agent-generated code in your pipeline, audit where review time actually goes: adding more AI output without expanding review capacity will widen the backlog, not shrink it.
Security debt compounds: Vikunja, Nginx UI, and Cloudflare tracing
A cluster of newly disclosed vulnerabilities in Vikunja and Nginx UI, combined with Cloudflare's new tracing layer, puts observability and access-control hygiene squarely on the week's to-do list.
Vikunja's default CORS config trusts all localhost origins, enabling credential theft
Vikunja ships with a permissive CORS policy that accepts any localhost port and sends credentials, letting any local page make authenticated requests on behalf of the user.
Why it matters: If you self-host Vikunja, patch or explicitly restrict cors.origins before any user on that machine visits an untrusted local app.
Cloudflare Traces enters open beta, turning the proxy layer into OpenTelemetry spans
Cloudflare Traces now automatically instruments Workers and security rules as OpenTelemetry spans, with volume-based pricing, giving teams end-to-end visibility across the edge without custom instrumentation.
Why it matters: For teams already running Cloudflare Workers, this is a low-friction way to get distributed tracing coverage across the edge tier without standing up a separate collector.
Build autonomous agents that intelligently navigate and analyze your data. Learn to develop agentic RAG systems using LlamaIndex, enabling powerful document Q&A and summarization.
Audit an AI agent workflow for unsafe autonomous actions
You are a senior AI safety reviewer. I will describe an AI agent workflow. For each step where the agent takes an action in the real world (sends a message, submits a form, calls an API, writes to a database), identify: (1) what could go wrong if the agent hallucinates or misunderstands context, (2) what the blast radius is (reversible vs. irreversible), and (3) what specific human-review gate or confirmation step should be inserted before that action executes. Be concrete and prioritize by severity. Here is the workflow: [PASTE YOUR WORKFLOW DESCRIPTION HERE]
Why it helps: Today's Philadelphia police tipline incident is a textbook case of an agent taking an irreversible real-world action without a human gate. Running this audit on your own pipelines before you ship is far cheaper than the alternative.
Before you ship it
The risk
AI agents connected to external submission systems (tiplines, ticketing systems, email, APIs) can act on hallucinated content before any human sees the output, turning a model error into an irreversible real-world event.
Do this
Insert a mandatory human-review step before any agent action that is external, irreversible, or has legal or safety consequences, and scope agent permissions so they cannot submit or publish without explicit approval.
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.
