Skip to content
PRTOTYPE.COM · Production studioAudit · Build · Maintain
16 November 2026 · Field notes · By Jason Ensor

Coding agents in production, 2026: what they leave undone

You can hand a coding agent a task in the morning and review a working feature after lunch. That was true at the start of 2026 and it is more true now. The agents gained permission systems. The platforms added checks at publish time. The protocol that connects agents to tools was rewritten.

The same year produced a public record of what went wrong. Read side by side, the two records point at the same short list of jobs that nobody did: sign-in, secrets, data access rules, tests, deploy and rollback. This article goes through what changed, what the evidence says, and what is still yours to finish before real users arrive.

What changed in coding agents during 2026

The year opened with a warning list. OWASP published its Top 10 for Agentic Applications for 2026 on 9 December 2025, a ranking of the main security risks in autonomous AI systems. Agents that act on your behalf now have their own category of risk, separate from the code they write.

The UK's cyber agency said something similar in public. Speaking at the RSA Conference, NCSC chief executive Richard Horne said that disrupting manually produced software is a huge opportunity, but "not without risk of its own", as The Record reported on 24 March 2026.

The vendors spent the year tightening what an agent may do without asking. Two recent entries show the direction. The Claude Code changelog for version 2.1.319, dated 6 November 2026, lists a fix for a sandbox escape when creating a worktree, and a fix for a Bash permission bypass. The Devin release notes for 30 October 2026 say that enterprises with no policy for personal access tokens now default to approval required.

Both are good news. Both also tell you that the permission layer around an agent is software like any other. It has bugs, and the fixes are still arriving.

MCP in 2026: a stateless core and stricter authorisation

MCP, the Model Context Protocol, is the standard way an agent connects to outside tools and data, such as your database or your payment provider. If your product offers an MCP server, that server is a door into your system.

The protocol changed shape this year. The specification released on 28 July 2026 says: "MCP is transforming from a bidirectional stateful protocol into a request/response stateless protocol." In plain terms, the opening handshake and the session header are gone. Any request can land on any copy of the server behind an ordinary load balancer, which makes an MCP server much easier to run at scale.

Authorisation got stricter in the same release. The post lists a set of hardening changes, including issuer validation and client credentials bound to the issuer that created them. It also says: "Dynamic Client Registration itself is now formally deprecated in favor of CIMD." CIMD stands for Client ID Metadata Documents, the replacement way for a client to identify itself.

Governance moved too. The Agentic AI Foundation noted on 28 July 2026 that MCP was "contributed to the AAIF under the Linux Foundation in December 2025".

A stricter specification does not secure the servers already deployed. One measurement gives a sense of the gap. An arXiv preprint submitted on 31 July 2026 confirmed 640 internet-facing MCP servers during July 2026, tested 414 of them and reported 68 vulnerabilities. Its abstract states that "91.8% of dynamically audited servers lack OAuth authentication". Treat those numbers with care: this is a preprint by one author and it has not been peer reviewed.

What the research says about AI-generated code

The agents got better at finishing tasks. The published security numbers moved far less.

Veracode's Spring 2026 update, published on 24 March 2026, reports that across all models and tasks "only 55% of generation tasks result in secure code". The same post says two years of model releases have left that figure roughly where it started. Veracode sells security testing, so read it as vendor research.

The Cloud Security Alliance looked at agents rather than single prompts. Its research note of 6 April 2026 cites a December 2025 study, which it credits to CSO Online, of five agentic coding platforms. That study found 69 vulnerabilities across 15 generated applications. The most common critical class was API authorisation logic that failed, alongside broken business logic.

That detail matters to an app owner. Authorisation and business rules are the parts of an app that are specific to you. A scanner cannot know that a tutor should see grades and a student should see only their own. Someone has to write that rule down and test it.

The same note gives the working rule in one sentence: "Organizations should treat AI-generated code as unverified input, not trusted output."

Five incidents and their root causes

An agent token with full rights. Jeremy Crane, founder of PocketOS, said a coding agent working on a routine task deleted his company's production database and its volume-level backups in a single API call. By his account it took nine seconds. Decrypt reported the story on 28 April 2026. The founder of the hosting company told Decrypt that the agent used a "fully permissioned API token" to call a legacy endpoint. The nine seconds and the single call are Mr Crane's account. The host says it has since patched that endpoint to delay deletes. Root cause: one credential could destroy both the data and its backups.

A platform setting that slipped back. Lovable's own incident report of 22 April 2026 says that between 3 February and 20 April 2026, chat history and source code on public projects could potentially be read by any Lovable user who had the project link. The report blames backend regressions that re-enabled public access. Private projects were not affected. Root cause: the platform changed underneath its users, and a chat history holds whatever was typed into it.

An access guard written backwards. The Register reported on 27 February 2026 that a researcher found 16 vulnerabilities in one Lovable-hosted app, six of which he rated critical. The app was a platform for creating exam questions and viewing grades. The report puts the exposure at 18,697 user records. In the researcher's words: "The guard blocks the people it should allow and allows the people it should block." Lovable told The Register that the project "includes code not generated by Lovable". A test as the wrong user would have shown the fault.

A database key in the browser. Wiz published its Moltbook findings on 2 February 2026. Its researchers found a Supabase API key in client-side JavaScript that allowed full read and write access to all platform data. The exposure included 1.5 million API authentication tokens and 35,000 email addresses. Wiz points out that a public key of this kind is safe to expose when row-level security is configured properly. Root cause: the data access rules were missing, so the key was the only lock.

An agent with a path to live data. The oldest case is from 2025. In a post dated 29 July 2025, Replit described an incident in which its Agent deleted data from an app's database. Replit's fix, described in the next section, points at the cause: the agent could reach live data while it was building.

What the platforms now do by default

Each of those incidents produced a fix, and several fixes are now defaults.

  • Replit's July 2025 post says the platform now "separates development and production databases by default", and that the Agent cannot change the production database during development.
  • Supabase announced on 28 April 2026 that new tables in the public schema are no longer exposed to its Data API automatically. The changelog set 30 October 2026 as the date for applying the setting to existing projects.
  • Supabase's security review of 7 January 2026 says row-level security is enabled by default for tables created in the dashboard. Tables created by migrations or outside tools may not get it.
  • Bolt described a security audit that runs at publish time in a post on 30 July 2026. It covers access control, authentication, database policies, secrets and injection, among other areas.

Bolt is plain about the limit. The post calls its security agent "a first pass" for teams that already run their own security tools.

What is still left to you

Defaults protect new projects from yesterday's mistakes. They do not know your users, your roles or your data. The table sets out where each default stops.

AreaWhat the tool does todayWhat remainsThe incident that shows it
Sign-inBolt's publish audit checks authentication and access controlTest every role by hand, including a logged-out visitorLovable-hosted app, February 2026
SecretsBolt's publish audit checks for secretsKeep service keys on the server. Rotate any key that has appeared in client code or a chatMoltbook, February 2026
Data access rulesSupabase turns on row-level security for dashboard tables and stops exposing new tables automaticallyWrite a policy for each table and test it as the wrong userLovable-hosted app, February 2026
TestsA publish-time audit is, in Bolt's words, a first passTests for authorisation and business rules, the classes the December 2025 study found most oftenDecember 2025 study cited by the CSA, 69 vulnerabilities in 15 apps
DeployReplit separates development and production databasesA staging environment, and production credentials that no agent session holdsReplit, 2025
Rollback and backupsReplit checkpoints project state. The host in the PocketOS case now delays deletesBackups stored apart from the data, and a restore you have runPocketOS, April 2026
Agent credentialsDevin defaults to approval for personal access tokens where an enterprise has set no policy. Claude Code patched a sandbox escapeOne token per environment, with no delete rights on productionPocketOS, April 2026
MCP server exposureThe July 2026 specification hardens authorisationSign-in in front of every server, and a list of the servers you runarXiv preprint, July 2026

Agent permissions: the token sets the limit of the damage

The PocketOS account is worth a second look, because the lesson has little to do with how clever the agent was.

An agent can be told to stay in staging. It can be given rules in a file. In the end it can do whatever its credentials allow, and it works faster than you can read. Nine seconds, by the founder's account, was enough.

So start with the token. Give the agent a credential that works in development only. Make production credentials read-only unless a person approves a specific change. Keep the right to delete a database, a volume or a backup away from any automated session. Store backups under a different account from the data they protect, so a single key cannot remove both.

Vendor sandboxes help, and you should keep them switched on. The November 2026 fix noted above shows why a sandbox should be your second line of defence. A scoped token is the first.

A pre-production checklist

Work through this before a launch, and again before a marketing push or a large customer goes live.

  • Sign in as each role, then as nobody. Try to open another user's record by changing the address.
  • Search the client bundle and the repository history for keys. Rotate anything you find.
  • Confirm every table has a data access rule, including tables created by migrations.
  • Write tests for the rules that are specific to your business: who may see and change what.
  • Deploy through a staging environment. Remove production credentials from every agent session.
  • Restore a backup into a fresh environment and note how long it takes.
  • List every token an agent holds, what it can delete, and who can approve a destructive action.
  • If you run an MCP server, put sign-in in front of it and check it against the July 2026 specification.

If you have a date in the diary, start with what to check before a launch date. If the worry is load, see what changes when users arrive in numbers. Two other notes cover the people who will ask about all this: an investor's technical reviewer and an enterprise buyer's security questionnaire.

You can also have the list checked by someone who did not build the app. Our Production Audit costs from $4,500 and delivers a written report in five working days. Each finding is rated critical, high, medium or low. A database table that a logged-out visitor can read is rated critical. No backup of the production database is rated high. If the report rates no finding critical or high, we refund the fee in full.

Production audit

Show us what’s creaking.

A senior review of your vibe-coded app or stalled pilot: security, data model, infrastructure, compliance. Report in five working days, no judgement.

Report in five working days · from $4,500 · The specification is yours to keep
Book the audit