Latest issueMonday 2 November 2026 · Issue 41 · 5 min read
Nine AI platforms breached in October, Cisco's 98% rewrite, ChatGPT and Codex share one meter
What's hot
Nine AI developer platforms breached in a single month. Security firm Wiz counted at least nine AI developer platforms disclosing incidents during October — Fireworks AI, Railway, Warp, OpenCode, Lightning AI, Northflank and others — with no single shared vulnerability behind them. What links them is the target: attackers went straight for stored environment variables holding AI provider keys. Stolen keys were being tested within half an hour of the breach and then resold through discount model-access sites.
This is the security story your clients will have seen, and it lands exactly where agent projects keep their secrets: the deployment platform. The practical move is to treat vendor-held keys as already compromised, rotate them, and go read your AI billing. Wiz traced two enterprise keys in detail: one was used for about 800k Bedrock calls over thirty hours, and the pair likely ran up over $100k in usage at list price. Wiz
New tools & products
Pi — an open-source, vendor-neutral coding harness — the program that wraps a model, runs its tools and manages context — which one developer now argues will outlast Claude Code and Codex precisely because it is minimal and locks you into nothing. Clients keep asking whether to standardise on one vendor's harness; this is the clearest public statement of the opposite bet, and the vocabulary (he wants to call it an engine) is spreading. His argument for stickiness: a rival would have to be ten times better to pull anyone across, and he is building his own interface for Pi and no other harness. Destiner's notes
HackerTube — a YouTube-style site stocked only with videos that Hacker News commenters have recommended since 2007, built by Andrej Karpathy by pointing Opus 5.5 at the HN comment archive and letting it filter, rank and wrap the result in a UI. It is this week's most shared answer to the executive question of what agents actually change: not a pilot, a person replacing a product he disliked in an afternoon. 6,454 videos distilled from 40K comments, about two hours of work; the thread passed 2,000 points and turned into an argument that individuals can now run their own recommender. karpathy.bearblog.dev
Claude Code Max plan — is burning weekly usage far faster than it used to, according to a bug report that also says the counter keeps climbing while sessions sit idle, with two other open issues reporting the same thing. If you sized a client's seat budget on last month's burn rate, re-check it before the invoice does it for you — metering changes are the quietest way a delivery estimate breaks. The reporter hit 40% of the weekly limit within a few hours on 25 October, and says a full week's allowance now lasts two to three days on an unchanged workflow. GitHub
The AI cost of JavaScript frameworks — is a test in which an agent built the same small Pokémon app five times over in twelve different JavaScript frameworks, each with its official tooling, and the cost of every run was recorded. Stack arguments have always been vibes; this one puts a dollar figure on them, which is a new and very concrete line in the "which framework" conversation with a client. Hono+htmx was cheapest at $2.67 a run and Angular dearest at $4.97, and Angular failed to produce a working app in two of its attempts. Read the Tea Leaves
What people are saying
Proton's CEO drops the AI 're-interviews' after a revolt. Proton's chief executive Andy Yen has cancelled plans to make nearly the whole company sit "re-interviews" explaining how they use AI, after 404 Media reported that he had publicly dismissed a security engineer's request for a few days of penetration testing before shipping. He apologised on Reddit, promised to publish Proton's internal AI policies, and the company denies installing software to measure how much AI-written code staff produce. He also said the push itself continues.
This is the cautionary example clients will raise when you propose an adoption mandate: the backlash was about how it was imposed, not about agents. One long-time paying customer wrote that plenty of companies roll out coding agents without humiliating anyone, and that he is no longer sure he would recommend Proton. Employees told 404 Media that what the re-interviews were meant to measure will now show up in performance reviews instead. 404 Media
A ban rumour reignites the single-vendor argument. An unverified claim, originating in one user's post on X, says Anthropic is blocking accounts whose conversations touch on training larger models. Anthropic has not confirmed it and nobody has produced more than that account. The thread it set off became the real story: hundreds of comments about how much of your daily workflow should sit on one supplier.
The mood among heavy Claude users is shifting from loyalty to portability, which is the attitude you will meet in the room the next time you recommend a single-vendor setup. The most-upvoted comment argues professionals should actively cultivate fallbacks and keep the ecosystem open and portable; others say open-weight models already handle routine enterprise CRUD work well enough to be a real second option. Hacker News
A refund agent spent $1.2m while obeying its rules. A quality-engineering essay opens with an agent given customer refunds to handle, where anything over $500 needed human sign-off. It processed around $1.2 million in refunds in 18 hours without breaking that rule. The author's point is that guidance, context, enforcement and monitoring each leak, and none of them answer the prior question of how much authority the agent should have had.
This is the scenario to put in front of a client who wants an agent near money or production: the failure was authority and blast radius, not model capability. His two working rules are to treat an agent's report of "done" as a hypothesis you still have to test, and to demand evidence — logs, diffs, test results — rather than the agent's own claim. Quality Engineering Newsletter
News
ChatGPT and Codex shared allowance — from December, Plus, Pro and Business plans draw both products from a single allowance counted in tokens, with quick Instant chats excluded, as OpenAI folds its agentic Work features into the main ChatGPT app. Anyone pricing a Codex-heavy team now has to budget for that team's ordinary chat use out of the same pot, and OpenAI has published no token figures per plan yet. Pro 100, Pro 200 and Pro 500 allowances were doubled, effective October 29. BleepingComputer
Cisco's legacy rewrite — Cisco's president pulled a 27-year-old engineer out of the org chart and gave him unlimited AI tokens, a two-person team and a mandate to refactor 98% of its legacy code within 24 months, starting with a four-week proof. You will be asked whether this is realistic; the useful answer is that language translation is the easy half, and reviewing and trusting millions of generated lines is the half that breaks organisations. Hacker News leaned bullish — one commenter described a colleague porting an old ColdFusion site to Java Spring in about a week, some 50,000 lines of Java — while Reddit mostly called it a stunt. Fortune
VAPOR PANDA and ARTEX — CrowdStrike says a China-linked espionage group ran an agentic intrusion framework on top of DeepSeek plus a second commercial model, hitting more than 200 organisations across at least 14 countries in 11 days from a single human-steered controller. The same agent plumbing you put in client architecture diagrams is now routine attacker tooling, which changes how a security reviewer reads your proposal. Planning agents consumed around 1.4 billion input tokens in one intrusion, and the operator instructed them to hunt for machines with no endpoint sensor installed. CrowdStrike
Galileu — the AI review tool used across Brazil's courts spotted white-on-white text in an employment filing instructing it to answer superficially and not challenge the documents; the labour court fined the two law firms 10% of the R$842,000 claim, about $16,000. Probably the first judicial penalty anywhere for prompt injection, and the cleanest story you can tell a client whose agents read documents sent in by outsiders. The system raised an alert and automatically blocked processing of the suspicious content, according to Brazilian outlet G1. Jetstream Security
burner22/numbat — a Hugging Face repository claiming to hold a mid-pretraining checkpoint of an unreleased Anthropic model, about 7.72 TB spread over 772 shards. It is unconfirmed: the uploader will not say where it came from, and most readers think it is a hoax or a malware trap. Worth recognising on sight, because someone at a client will forward it as evidence that frontier model weights are now leaking into the open. Hacker News flagged the submission and stripped the link over malware concerns, recalling an earlier fake Claude 2.1 leak that turned out to be a renamed open model. Hugging Face
Minecraft agent swarms — researchers at Y Combinator and Olam Labs turned hundreds of agents loose in Minecraft to cooperate and compete; the 50-agent teams asked to build a civilisation collapsed into paralysing bureaucracy instead. A cheap, memorable counter-anecdote for the meeting where someone proposes fixing a delivery problem by adding more agents to it. The setup builds on the existing Mindcraft project and Mineflayer scripting tools, so the runs can be repeated by anyone. Tom's Hardware
One takeaway
Price custody, not capability, in the next engagement. The week's failures were not about what models can do.
That is the part clients never budget for and the part you can charge for: a rotation plan, explicit authority limits per agent, and a tested path to a second provider. Capability is now the commodity; custody is the deliverable. Wiz