The AI Teardown· August 8, 2026

Samsung Support Pasted Its Own Prompt Into the Customer Chat

A Samsung support agent, live in a customer chat, pasted the raw ChatGPT prompt they were feeding the model straight into the conversation.

No jailbreak, no leak tool, just a copy-paste between a hidden window and a customer. It was the fastest-moving post on the board today.

The Human Clipboard Is the Leak

It is an image post, so the title does the work, and the title is enough: the agent pasted its prompt into the support window.

The top comment delivered the verdict and added nothing, which is how the best ones work: “AI is trying to take our jobs of copying text from ChatGPT.” A second: “im surprised its not a chatbot in the first place”.

The lesson is not that the model failed. It is that a support pipeline with a human clipboard in the middle is structurally leaky.

Either the bot is fully instrumented with guardrails on what it can emit, or model output never touches a human relay. There is no third setting where the seam holds.

And a leaked system prompt is not just embarrassing. It is free reconnaissance for the next prompt-injection attempt.

Five Token-Saving Tools, and Nobody Saved 60%

The most builder-useful post of the day is buried at 18 upvotes. OP ran 48 Django questions from SWE-bench across five token-saving tools plus a no-tools baseline, same agent, same repo commit, 261 runs in total.

The framing they supply is the July record: JetBrains reran two headline claims and found Caveman claimed 65 percent but measured 8.5, while RTK claimed 60 to 90 percent and came out slightly more expensive than using nothing.

Measured against an 1,828-token baseline on Codex: repowise cut output 31.6 percent, CodeGraph 24.4, Serena 14.8 at a cost of 10.1 tool calls per question, Graphify 8.9, and code-review-graph 6.0.

OP’s own summary: “Nobody saved 60%.” and “The best result was about 32% of the agent’s output tokens”.

The tell is that OP works on repowise, which tops the chart, though they also shipped the harness, raw data and preregistration. The real envelope is 15 to 32 percent, and a benchmark without a bill is still marketing.

Claude’s WebFetch Hands the Model a Summary, and It Invents

OP had Opus 5 research memory architecture and kept getting clean-looking stats and quotes that were wrong. Asked directly, the model said it never read the papers.

WebFetch hands it a summary from a smaller, cheaper model. That model “compresses things, guesses, and apparently sometimes invents details”, and one fabricated citation named a framework that was not in the paper.

The fix was Sonnet subagents with a no-WebFetch rule, curling the page and reading the actual text. It caught 17 errors across roughly 30 papers, two of them “basically been reported backwards”.

The doctrine is plain. A summarizer sitting between the model and the source turns citation-grade research into confident fiction, and reframes a lot of hallucination complaints as pre-mangled input.

Stack Overflow Fell From 207,000 Questions a Month to 1,400

A chart post marks the peak at 207,000 questions in March 2014 and July 2026 at 1,400. Call it a 99.3 percent collapse.

SO was the quick-answer and training-data backbone for a generation of developers, and net-new Q&A has effectively stopped, even though the models that replaced it learned from that exact corpus.

The comments named the human cause. One remembered a culture that treated beginners as if you needed to be a Jedi to ask a question, which survived pre-LLM but “Now it’s toast and the LLMs are already better than most coders”.

The epitaph wrote itself, mocking SO’s own reflex: “Your post has been marked as duplicate and downvoted to satisfy other users egos”.

The moat SO trusted, its pile of accumulated answers, became the training set that replaced it. Nobody in the thread answers where the next model learns about APIs that postdate the frozen corpus.

Off the Bench

Anthropic shipped cross-session messaging in Claude Code so top-level sessions can DM each other, and the room’s shrug was that it had rolled the same thing with files and tmux for months. Passing messages was never the hard part, “preventing agents from overwriting each other’s work” is.

A week on two $200 subscriptions produced the day’s best line, “Sol writes like a lawyer, but Fable bills like one.” The keeper under the snark is the crew: Fable as architect, Codex on Sol as the builder that “coded like it owed somebody money”, and Sol Pro as the adversarial reviewer.

Anthropic’s OpenRouter traffic share is sliding, and the 142-comment thread converged on one cause, price. One commenter put DeepSeek V4 Flash at $0.14/$0.28 per million against Opus at $5/$25, concluding “you have to be kind of dumb to pay for Claude outside of their subsidized plans”.

OP replaced a complex Reddit research agent with a single Codex skill, dropping the agent loop and orchestration framework for a workflow file plus small deterministic helpers where reproducibility matters. The blunt thread consensus, “90% of people on this sub are over engineering stuff codex and Claude can do”.

The one job worth doing this week, keyed to Samsung: grep your own support and agent logs for your system-prompt strings, because if a bored human relay can paste them to a customer, an attacker can pull them just as easily.

By the numbers

2,772 upvotes and a velocity near 390 an hour for the Samsung prompt leak, the day’s fastest post.

32 percent: the best token saving measured across five tools, against claims of 60 to 90.

17 fabrication errors caught across roughly 30 papers once WebFetch was removed.

207,000 to 1,400: Stack Overflow questions per month, March 2014 to July 2026.

$0.14/$0.28 versus $5/$25: DeepSeek V4 Flash and Opus, per million tokens in and out.