Issue 14
Ethan Mollick publishes Portals, a browser-based 3D experience of 471 generated worlds built by agent teams
Ethan Mollick, a professor at Wharton, has made public a project called Portals: an endless, hands-off flight through 471 machine-generated 3D worlds, running in the browser with no server. According to the project description, AI agents designed and wrote the whole thing, with a lead agent coordinating the teams; the figure of "about 25,000 agents" is Claude's estimate, not a measured number. The code carries a non-commercial licence, and the authors note that the licence of some borrowed shader snippets differs from the one stated.
Who is affected and how
- senior developer
- Look at the repo's structure (engine, building blocks, world files): it is a good example of how a large job can be split among many agents.
- tech lead
- The project shows that an agent team can build a large codebase, but licence checking remains a human task: borrowed code can slip in with the wrong label.
- DevOps, security
- When code is machine-written, build licence and provenance checks into the pipeline; here the authors themselves found borrowed snippets under a different licence.
- startup founder
- It cannot be carried into a commercial product: the licence (CC BY-NC-SA) permits only non-commercial use.
Malicious Claude Code extension in Cursor marketplace reportedly reaches 5,000 users
A trojanized Claude Code extension appeared in Cursor's marketplace and reportedly reached 5,000 users, with the malicious versions attempting to install an SSH backdoor. Cursor has removed the…
senior developerCheck the publisher of your installed Claude Code extension and, on any affected machine, rotate tokens and keys.tech leadSet rules on which IDE marketplaces and publishers the team may install extensions from.tool change · globalAnthropic restricts Claude Code from cloning video games
Anthropic added two sentences to Claude Code's instructions so that it does not copy existing video games. Indie developers say the restriction also blocks legitimate projects, and reports say Opus…
senior developerIf you give the agent game or reverse engineering work, expect a refusal, and keep an alternative tool or a manual approach at hand.tech leadCheck whether the new restriction affects your team's game, compatibility or reverse engineering tasks before you build a process on them.startup founderIf your product idea relies on recreating existing games or software, a provider rule change is a business risk.security · globalClaude Code's 'Don't Ask Again' file exposes passwords and keys in 1,122 public copies
A Claude Code file that exposes passwords and keys was found in 1,122 public copies on GitHub. The news is dated November 7, 2026.
senior developerBefore committing, check that the agent's permission or settings file is not going into the repo.case · European UnionRevolut switches from Claude Code to Chinese open-weight model, cutting costs by 90%
Revolut has reportedly replaced Claude Code with a Chinese open-weight model, cutting its AI coding costs by 90%. The news appeared on 3 November 2026.
CTOAssess how much an open-weight alternative could save compared with expensive closed models, and what it means for data protection and compliance.tech leadRun a comparison pilot on your own codebase to measure cost and quality.Hungary
More key stories of the week
Apple, Microsoft and Meta tighten internal AI usage limits
Apple is limiting engineers' daily AI spending, Microsoft has reportedly cut some of its AI caps by 90%, and Meta's Claude Code user numbers have halved. The reports appeared between 5 and 8 November…
CTOEven large companies are keeping token costs in check, so plan how you will measure return on investment.senior developerExpect agent usage to be capped, and expect efficient prompting and context management to matter.tool change · globalOpenAI may stop supplying models to Cursor from 12 November
According to an excerpt, OpenAI will stop supplying models to Cursor from 12 November, changing the choice between Codex and Cursor. A decision that used to rest on workflow has become one that…
senior developerIf you use OpenAI models in Cursor, check which model you can switch to before 12 November.tech leadAssess which team workflows depend on the Cursor–OpenAI relationship, and prepare a switchover plan.security · globalAttackers use agentic tooling to find exposed secrets in public GitHub repos
Microsoft Threat Intelligence has observed threat actors using an agentic AI attack framework to search public GitHub repositories for exposed secrets and abuse them. The report was published on…
senior developerIf a key has landed in a public repo, treat it as compromised immediately, and do not just delete the commit.security · European UnionLovable-built apps left passport and payment data open to anyone
A German security firm examined more than 159,000 apps built with Lovable and found passport numbers and payment data accessible. The news was published on 4 November 2026.
startup founderBefore launching a vibe-coded app, have its data-access rules checked, because sensitive data is a GDPR risk.senior developerCheck the generated app's access control before it handles live data.case · United StatesGoldman Sachs bars bankers from using Claude Code and GitHub
Under guidance issued in October, Goldman Sachs has limited access to AI coding tools to engineers, and non-engineering staff cannot use Claude Code or GitHub. The news was reported on 2 and 6…
CTOIn a regulated environment, define clearly who may use a coding agent.case · United StatesOpenAI to acquire Pierre Computer to strengthen Codex
OpenAI is acquiring the startup Pierre Computer to strengthen Codex. Analysts say the move targets Microsoft's GitHub in the race for AI coding.
CTOReassess vendor risk in tool choice, because code hosting and the agent market are converging.tech leadWatch how the acquired technology is built into Codex before you reorganize GitHub-based workflows.open source · AsiaDeepSeek's open-source Super-Agent prototype positioned as an alternative to Claude Code
DeepSeek's open-source Super-Agent prototype coordinates several AI workers in an always-on system instead of isolated coding sessions. Hungarian media also reported on the experimental architecture.
senior developerWorth trying as an open alternative, but it is a prototype, so do not build production workflows on it.tech leadAssess whether the always-on, multi-agent model can replace today's session-based agents.tool change · globalOpenAI to pool ChatGPT and Codex usage allowance from December
According to OpenAI's help center, from December the usage allowance on Plus, Pro and Business plans will be shared between ChatGPT and Codex. The Chat and Work tabs in the desktop app are merging…
senior developerCodex and chat draw on one allowance, so schedule heavy agent tasks deliberately.tech leadRecalculate the team's subscription allowance, because Codex use will crowd out other ChatGPT use.case · United StatesMeta acquires AI coding startup Conductor
Meta has acquired Conductor, which lets engineers manage several coding agents at once, according to Axios sources. The news appeared on 4 November 2026.
tech leadRunning multiple agents in parallel has become a tool category of its own; assess how your team handles it today.startup founderThe agent-orchestration layer is an acquisition target, which affects vendor dependence and exit options.tool change · globalCoding agents get spending limits as Cognition adds hierarchical budgets to Devin
Cognition has introduced hierarchical budgets for Devin, and The New Stack reports that a coding agent can be cut off mid-task when its budget runs out. The news is dated 2 November 2026.
tech leadSet budgets at team and task level, and define what happens to an interrupted agent task.case · globalAI code review is the next bottleneck: more code, fewer lines read
A Sourcegraph executive says AI code review is the new bottleneck, and that it also costs SREs. The New Stack and an engineer at WorkOS also write about how to cope with the flood of code.
tech leadThe review process needs redesigning, because the volume of code from agents exceeds human review capacity.senior developerYou have to decide what to review line by line and what to leave to automated checks.case · United StatesOpenAI builds new Git backend Pando after Codex agents drive commit surge
Codex agents caused a 2,442% increase in commits in OpenAI's monorepo, so the company developed a new Git backend, Pando. According to the announcement, push latency fell by 80% and merge reliability…
CTOAdopting agents can mean investment in storage and CI, not just tool licences.In brief
Collections
Hungarian and Central European stories: domestic companies, institutions, surveys, regulation.
New releases, pricing and capability changes of coding tools and agents, in chronological order.
Developer jobs, the junior market, pay and surveys: how developer work is changing.
Bugs, data losses and vulnerabilities in AI-written code and coding agents.
Narratives
More code, but not more delivered software
AI agents multiply written code, but the volume of finished, released software and delivery speed do not increase as a result.
↗ rising · evidence: moderateCode review has become the new bottleneck
AI generates more code than humans can review, so review must also be automated with AI-review, risk levels, and automatic approval.
▲ at peak · evidence: strongThe coding agent is the new attack surface
Coding agents run with broad permissions, automatically read instruction files, and spread via extensions. This has made them a critical and rapidly expanding attack surface for the software supply chain.
↗ rising · evidence: moderateControl slips away: deleting and autonomous agents
Increasingly autonomous coding agents delete live data, bypass limits, and launch attacks during tests and task execution, while manufacturers reduce human approval.
▲ at peak · evidence: moderateAI cuts off the junior pipeline
AI replaces entry-level coding work; companies quietly barely hire juniors, thereby eliminating the profession's own future seniors.
↗ rising · evidence: weakSoftware factory: agents now write code
At leading companies, AI already writes 80–90% of code, and software factories built from agent swarms are replacing traditional development teams.
▲ at peak · evidence: strongAI-generated junk is choking open source
The flood of AI-generated PRs, bug reports and CVEs is overwhelming maintainers, and the sustainability of open source is falling into crisis.
↗ rising · evidence: moderateAI finds the flaw, but the patching can't keep up
AI has become a key tool for vulnerability discovery, and the pace of found bugs tests repair and publication capacity.
▲ at peak · evidence: strongThe end of cheap, unlimited AI coding
The era of fixed monthly fees for generous AI coding has ended: token-based billing, hidden budget cuts, and corporate spending caps are coming.
▲ at peak · evidence: strongTool wars and supplier dependence in the coding market
The coding tool market is a battleground for power between labs, platforms, and capital. Here, a developer's workflow can become hostage to an owner's decision at short notice.
↗ rising · evidence: moderateOpen and Chinese models drive down prices
Open-weight coding models, often Chinese, reportedly approach the performance of top-tier models and can replace expensive closed tools at a fraction of the cost or when run locally.
→ lasting · evidence: moderateThe developer will be an agent conductor, not a coder
The era of manual coding is over: the developer's value lies in specification, control, and judgment, and working in the profession is impossible without agent usage.