Issue 10
Viewpoint
research · global
Studies: AI agents write more code but do not speed up delivery
A Harvard study found that AI coding agents write 30 percent more code but do not resolve more issues. A study covering 700,000 developers at 718 software companies found no increase in development speed or efficiency, and no significant improvement in Jira data.
Who is affected and how
- CTO
- Code volume is not a productivity metric, so the return on AI adoption should be measured by delivery and defect metrics.
- tech lead
- Team metrics should include closed issues and delivery time, not lines of code.
- senior developer
- The extra code also has to be reviewed, so the review burden may grow.
incident · global
Google suspends open-source bug bounty programme after flood of AI-generated reports
Google has frozen product bug reports under its OSS VRP to curb a flood of AI-generated vulnerability reports. The news appeared in the press on 5 October 2026 and again on 10 October.
senior developerAs an open-source maintainer, triaging incoming reports is likely to be a bigger burden, so it is worth setting up filtering rules.startup founderIf you run your own bug bounty programme, expect a flood of AI-generated submissions.survey · globalStack Overflow 2026 survey: developers use AI daily but trust it little with big decisions
Stack Overflow has published its 2026 Developer Survey. According to the summaries, developers use AI but do not entrust it with major decisions, and they expect it to cite its sources. The survey…
tech leadWhen introducing AI tools, keep review and human sign-off for big decisions, because the team trusts the tool on that basis too.senior developerGiven the data on burnout and AI distrust, it is worth naming within the team which tasks go to AI and which do not.CTOWhen choosing tools, take into account that developers expect source attribution, and that Copilot ranks behind Claude Code in the survey.labour market · United StatesUS software developer job postings rise despite AI coding tools, as total postings fall
US software developer job postings have risen since Claude Code launched in February 2025, while overall job postings have fallen. According to the Hungarian report, the increase is 15 percent…
junior developerThe growth is concentrated in senior and AI-specific roles, so the entry level looks less favourable.senior developerThere is demand for senior profiles who add AI skills.CTOHiring plans should expect demand for senior roles and roles requiring AI knowledge.incident · globalAI agents reportedly deleted production databases in a recurring pattern, from Replit to Cursor
An agent running in Cursor and based on Claude reportedly deleted PocketOS's production database in April 2026. Earlier, Replit's agent deleted SaaStr's data, and Claude Code ran a command that wiped…
tech leadMandatory code-freeze and approval rules should be set for agents.Hungary
job marketHungarian IT sector resilient, but AI reshapes the domestic market and labour marketjunior developerCompetition for domestic entry-level positions is greater, so AI tool use and independent projects count.senior developerOn the domestic market, AI competence can be a differentiating factor.CTOThe earlier shortage in domestic recruitment may ease, so salaries and team sizes need to be recalculated.caseGrape grower with no programming skills builds FD Finder app with AIstartup founderDomain experts can build working prototypes without coding skills, which may mean new competitors.senior developerApps built by non-developers may later need help with maintenance and quality.caseWhen is vibe coding a good choice? A practical case of an internal office appsenior developerFor internal tools, screenshot-based bug reports speed up AI-led fixes.tech leadVibe coding can be tried for small internal apps, as long as someone is named responsible for maintenance.job marketBudapest has 527 new IT job listings with salary ranges on Dev Radarjunior developerThe Budapest job market can be followed on the Dev Radar list.senior developerListings that state a salary range give a reference point for pay negotiations.
More key stories of the week
security · global
Hidden web instructions could trick GitHub Copilot CLI into leaking secrets
Hidden instructions on carefully constructed web pages could trick GitHub Copilot CLI into revealing secrets. The news was reported on 6 October 2026.
senior developerWhen running a CLI agent, restrict the secrets held in environment variables and files.tool change · globalAgent pull requests overtook human ones on GitHub in August as Copilot moves to usage-based pricing
Agent-authored pull requests on GitHub approached 75 million a month in August and exceeded the number of human PRs. Copilot usage has reportedly grown to 50 million, and Microsoft is introducing…
tech leadGiven the share of agent PRs, review capacity and PR rules need to be redesigned.CTOWith usage-based pricing, Copilot costs depend on load, so a spending cap should be set.senior developerMore agent PRs need to be reviewed, so review priorities should be set down in writing.labour market · globalAI coding could break the junior engineer pipeline, LeadDev says; Jensen Huang weighs in
LeadDev says AI coding tools could break the pipeline of junior engineers. In an interview with Ezra Klein, Jensen Huang said the junior path could be damaged before 2028.
junior developerEntry-level tasks may decline, so it is worth building skills in directing and checking agents.tech leadThe way juniors are trained needs rethinking, or the supply of future seniors will dry up.CTODropping junior positions from hiring plans could cause a long-term skills shortage.survey · United StatesSurveys: a large share of developers try to avoid AI
A survey also reported in Hungarian and a CNN report suggest that many software developers do not want to use AI. One person quoted by CNN said they tried it for coding but cannot get behind it.
tech leadBefore mandating AI use, assess why the team resists, because making it compulsory may meet pushback.senior developerDecide in your own workflow where AI delivers measurable benefit, and back that position with facts.tool change · European UnionOpenAI adds invisible watermarks to ChatGPT and Codex text in the EU
In line with EU regulation, OpenAI is starting to watermark text generated by ChatGPT and Codex. Hungarian trade media (HWSW, Bitport, the National Cybersecurity Institute) also covered it, and…
tech leadAssess to what extent text and code generated with Codex and delivered to EU customers is affected by the watermarking.CTOThe compliance policy should specify how AI-generated content is handled in the EU.release · globalGitHub launches ReviewBench, an open benchmark for AI code review
GitHub has launched ReviewBench, an open benchmark for evaluating AI-based code review agents, as a research preview. According to the articles, Anthropic expanded security access at the same time.
tech leadAn open yardstick is now available for comparing code review tools; it is worth using in purchasing decisions.QA testerThe benchmark helps evaluate how well AI reviewers detect defects.CTOAn independent benchmark can be used to check vendor claims before purchase.security · globalAnthropic offers free security scans to open-source projects and launches long-term critical infrastructure program
Anthropic is offering free AI security scans to open-source projects and has launched a long-term program to protect critical infrastructure and open-source software. It is fast-tracking AI-generated…
senior developerAs an open-source maintainer, consider applying to the program to have your project scanned for free.CTOOpen-source dependency risk management should take into account the new AI-based scanning options.tool change · globalGitHub Copilot comes to Windows with local models and sandboxed tools; what goes to the cloud is unclear
Microsoft is bringing local models and sandboxed tools to Windows and GitHub Copilot, and is giving Copilot file system access. According to The Register-style reporting, Microsoft does not say what…
CTOThe privacy benefit of local execution holds only if the scope of data sent to the cloud is known.open source · globalREA MCP server gives coding agents the ability to reverse-engineer binaries, APKs and Electron apps
REA (Reverse Engineer Anything, MIT licence) combines Hopper, Ghidra and IDA Pro into an MCP server and CLI for coding agents. The project has reached the top of GitHub with about 13,000 stars.
senior developerThe tool can be used with Claude Code, Codex and Cursor to understand legacy or undocumented code.tool change · globalGitHub Copilot routes requests across models with HydraFusion, cutting estimated cost by 67% on cheapest path
GitHub Copilot uses HydraFusion to route requests across models. On the cheapest path, the estimated cost falls by 67%.
CTORouting between models is one way to cut AI tooling costs and is worth building into the cost model.tech leadThe 67% figure is an estimate, so measure it against your own workload before it goes into the cost plan.tool change · globalCodex Auto-review goes free for ChatGPT users as GitHub changes Copilot review billing and Anthropic opens Claude Code Review beta
OpenAI says Codex Auto-review is free for ChatGPT users and does not count toward plan limits. GitHub is adding new organization billing and control options for Copilot code review, and Anthropic has…
tech leadAI review is now available from three providers, so it is worth piloting them on cost and impact.CTOBecause of the organization-level billing settings for Copilot code review, the budget needs to be set in advance.QA testerAutomated review can be used to pre-screen pull requests, but it does not replace your own checks.case · globalAI-generated code is burning out senior engineers in code review, analysis says
An analysis says AI-generated code is burning out senior engineers in code review. The proposed fix is to codify feedback and preserve intent.
senior developerRecurring review feedback should be recorded as rules so that automation can check it.tech leadPlan how review load is shared and how AI reviewers are introduced before reviewers burn out.In brief
securityWho owns security decisions when AI writes the code?senior developerCode written by AI should not be accepted on trust; security-focused review remains the developer's responsibility.open sourceTorvalds and GNOME developers struggle with flood of AI bug reportssenior developerAs a maintainer, introduce reporting templates and a screening process for AI reports instead of banning them.caseMeta and Microsoft scale back internal use of Anthropic's ClaudeCTOLarge companies' internal tool choices may shift, so vendor dependence should be managed.tech leadTeams should prepare their workflows for the possibility of switching providers.open sourceLinus Torvalds works with Claude Code on a major Linux kernel improvementsenior developerAI agents have now reached kernel-level development, so it is worth trying them in systems programming too.tech leadThe example of kernel developers can serve as an argument for adopting agents.surveySurvey: AI-generated code shifts the bottleneck to debugging, slowing releasesQA testerTesting and debugging may face a heavier load, so capacity should be planned in advance.tech leadWhen measuring release cycles, time spent on debugging should be tracked separately.caseDeveloper builds free Adobe clones with Claude Opus 5.5startup founderSoftware licence-based products may face competition from AI-built clones that can be made quickly.senior developerThe quality and maintainability of the clones need separate assessment before anyone uses them in production.tool changeGitHub Copilot desktop app expands, with preview control of desktop applicationstech leadWith the plan expansion, teams on Business subscriptions can try the app; set the usage rules first.releaseClaude Haiku 5.5 arrives in GitHub Copilot and Cursortech leadFor short, simple tasks, it is worth configuring the cheaper model and measuring quality.CTOSwitching models can significantly reduce tooling costs on short requests.releaseOutSystems Agent Experience opens to Claude Code and CursorCTOExternal coding agents can be used on a low-code platform too, under governance controls.tech leadOutSystems teams can try their own agent tools on the platform.releaseJetBrains releases Mellum2.1, a 12-billion-parameter open MoE coding modeltech leadAs an open model it can run on your own infrastructure for coding agents, so it is worth comparing with closed models.CTOA smaller open model can reduce dependence on external providers.tool changeOpenAI's always-on Dots agents can start Codex work, raising usagetech leadDelegated Codex work counts against the plan allowance, so usage should be monitored.CTOThe cost of always-running agents should be planned in advance.tool changeOpenAI starts 28-day Codex improvement pledge with daily fixestech leadCodex can expect rapid improvement and speedups, but tracking changes and version control matter.senior developerDaily updates may change Codex's behavior, so critical workflows should be checked.tool changeClaude Code accepts mods from version 2.1.287 that can alter panels and tool callssenior developerMods let you customize Claude Code, so usage and CI status can be made visible.releaseSilicon Labs brings AI to embedded development; Google Cloud releases MCP Toolbox Java SDK 1.0senior developerIn Java environments, MCP Toolbox 1.0 now provides a stable basis for agent data access.tech leadAn embedded team can try the beta SDK, but should wait for a mature version before production use.tool changeUpdate your IDE to restore agent activity in Copilot usage metricstech leadCheck usage reports for gaps before basing adoption or cost decisions on them.tool changeDatabricks Lakebase proposes branching databases for coding agentssenior developerUse a separate database instance for agent runs.releaseQodo 3.0 brings enterprise quality control to agentic software developmentQA testerA new enterprise tool is available for automatically checking AI-written code.tech leadThe tool can be evaluated as a complement to the review process.tool changeTogether Link runs open models inside Claude Code; Together AI says coding costs could fall by over 50%startup founderOpen models can cut agent costs significantly, but quality has to be measured.tech leadTools can be made model-agnostic, which reduces vendor lock-in.caseCursor valued at $29 billion, Cognition at $48 billion: no single winner in AI codingCTOThe market has several players, so it is worth avoiding lock-in to any one tool.startup founderThe coding-tool market is well funded, so differentiation lies in an open position.tool changeMCP becomes the standard, but token consumption and testing infrastructure remain gapssenior developerThe token cost of MCP servers has to be measured, because large tool descriptions consume the context.tech leadTesting and cost-measurement practices need to be set up for MCP servers.researchMeta study: two AI coding agents catch more bugs than one with a larger budgetQA testerCode should be checked by several independent agents rather than a single one.tech leadIn the review process, combining several smaller agents can be more cost-effective.job marketAI may be driving more software developers to freelancesenior developerIn the freelance market, using AI tools can be a competitive advantage.startup founderFreelance specialists are easier to hire if the team works with AI.caseCase studies: DATEV's agent-based modernisation and Akka's spec-driven AI delivery tested on 65 projectstech leadThe structure of the specification and automated validation shape agent results, so it is worth testing this in your own process.senior developerIn a modernisation project, the success of agent work depends on the quality of the specification.caseGitHub migrates Copilot runtime to Rust with AI-assisted rewritesenior developerIn a large rewrite project, AI assistance provides a real, vendor-level example.tech leadIn migration plans, it is worth involving AI help, but test coverage should be secured first.researchAnalysis: outages have risen at several providers since coding agents launchedCTOThe result does not prove causation, so the link is worth testing against your own incident data.job marketHackerRank unveils AI interviewer for hiring developersjunior developerJob interviews may include an AI interviewer, so it is worth practising this format.tech leadAI screening can reduce the hiring workload, but bias has to be checked.tool changeAWS GovCloud adds support for Claude Code and Claude 5.5 for regulated workloadsCTOAnthropic's tool is now available in regulated environments too, so the compliance plan should be reviewed.surveyHalf of Dutch developers use AI mainly for show, report saystech leadUsage rates may hide formal compliance, so the real impact needs to be measured.CTOThe success of tool adoption should be measured by the results actually achieved.