Sources

The AI SDLC Research Library

Every study, survey, and standard the AI SDLC Playbook cites, distilled to the one finding that matters from each and linked back to its primary source. If a number on this site surprises you, trace it here.

Last reviewed: September 2026

Before You Quote a Number

BaxBench's 62% counts incorrect-or-insecure, not insecure alone. The figure most often repeated as 'AI code has vulnerabilities' actually measures the share of solutions from the best model at the time that were either wrong or contained a security vulnerability. The 2026 leaderboard sits near 50%.

METR's 19% slower applies to a narrow population, not developers in general. The randomized controlled trial measured 16 experienced open-source maintainers working in their own repositories. METR's February 2026 update called the extended result 'very weak evidence' and is redesigning the study.

Vendor self-reports of AI-authored code use different definitions. When a vendor states a percentage of code 'authored' by AI, that can mean assisted or autonomous, written or merged, depending on which vendor and which report. Treat every such figure as a self-report, not an independent audit.

Jump to a Group

Productivity and Adoption

Faros AI, "The AI Productivity Paradox"

July 2025

Telemetry from 10,000+ developers across 1,255 teams: tasks completed per developer +21%, PRs merged per developer +98%, PR review time +91%, average PR size +154%. Company-level throughput and DORA metrics showed no significant correlation with AI adoption.

Primary source

Faros AI, "AI Engineering Report 2026: The Acceleration Whiplash"

May 2026

22,000 developers, 4,000 teams: bugs per developer +54% (up from +9% a year earlier), incidents per PR +242.7%, median PR review time 5x, code churn +861%, deployments per week -11.7%, PRs merged with no human review +31.3%.

Primary source

DORA, "Accelerate State of DevOps Report 2024"

2024

About 3,000 respondents: a 25% increase in AI adoption was associated with a 1.5% decrease in delivery throughput and a 7.2% decrease in delivery stability.

Primary source

DORA, "State of AI-assisted Software Development 2025"

09-23-2025

Nearly 5,000 respondents: 90% use AI at work, more than 80% say productivity is up, 59% say code quality is up. AI adoption is positively associated with throughput but still negatively associated with stability.

Primary source

DORA, "AI Capabilities Model"

2025

Names seven organizational capabilities that turn AI adoption into performance, including a clear AI stance, AI-accessible internal data, and working in small batches. AI amplifies whichever capabilities an organization already has.

Primary source

DORA, "ROI of AI-assisted Software Development"

2026

Introduces the J-curve, where productivity dips before it rises, and the verification tax: the review and validation cost that AI adoption shifts onto humans.

Primary source

METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity"

07-10-2025

Randomized controlled trial, 16 experienced open-source maintainers, 246 tasks on their own repos: with AI tools they were 19% slower, despite forecasting 24% faster and believing afterward they had been 20% faster.

Primary source

METR, uplift study update

02-24-2026

57 developers, 800+ tasks with late-2025 tools: the original cohort still estimated at -18% (confidence interval -38% to +9%), a new cohort at -4%. METR called it "very weak evidence" and is redesigning the study.

Primary source

GitHub, "Research: quantifying GitHub Copilot's impact on developer productivity and happiness"

09-07-2022

GitHub's original productivity-and-happiness study, published as Copilot moved toward general availability. It is the baseline research the much larger 2025 and 2026 adoption studies in this library build on.

Primary source

GitHub, "GitHub Copilot for Business is now available"

02-14-2023

Marked Copilot's move from an individual tool to a licensable enterprise product, the start of the seat-based procurement pattern this library's cost model page addresses.

Primary source

GitHub, Octoverse 2025

10-28-2025

The Copilot coding agent (renamed cloud agent in 04-2026) authored over one million PRs between May and September 2025.

Primary source

GitHub, "60 million Copilot code reviews and counting"

03-05-2026

60 million Copilot code reviews, more than one in five code reviews on GitHub; 71% surface actionable feedback, at 5.1 comments per review on average.

Primary source

McKinsey, "The State of AI"

Editions March 2025, November 2025, August 2026

November 2025: 88% of organizations use AI in at least one business function, yet only about 6% qualify as high performers. August 2026: 80% report individual productivity gains, but only 37% report any enterprise EBIT impact.

Primary source

Stack Overflow Developer Survey 2025, AI section

2025

49,000+ respondents: 84% use or plan to use AI tools, but only 33% trust its accuracy and 46% distrust it. The top frustration, at 66%, is "almost right, but not quite."

Primary source

Sonar, "State of Code Developer Survey Report"

January 2026

1,100+ developers: 42% of committed code is AI-generated or assisted, 96% do not fully trust it, and only 48% always verify AI code before committing.

Primary source

Anthropic Economic Index reports

January, March, and June 2026

The self-report series behind the claim that more than 80% of code merged to production in May 2026 was authored by Claude. Treat it, like every vendor self-report in this library, as using its own definition of "authored."

Primary source

Quality and Security

CodeRabbit, "State of AI vs Human Code Generation Report"

12-17-2025

470 open-source PRs, 320 AI co-authored and 150 human-only: issues per PR were 1.7x higher for AI co-authored work (10.83 vs 6.45), and security vulnerabilities were 1.57x more common (XSS alone 2.74x).

Primary source

Qodo, "2025 State of AI Code Quality"

06-12-2025

609 developers: 65% say AI misses relevant context and 76% do not fully trust generated code. Among respondents reporting productivity gains, 81% using AI code review saw quality improve, but Qodo's own product telemetry found 80% of PRs enabled for AI review get no human comment, and 17% contain a high-severity issue.

Primary source

Qodo, "Reviewing AI-Generated Code" (Gatepoint survey)

07-03-2026, 100 engineering directors and VPs

94% use AI coding tools and 39% have standardized, but only 12% are very confident in AI code; 62% still rely on manual peer review as the primary control, and 50% of standardized adopters cite review bottlenecks.

Primary source

GitClear, "AI Copilot Code Quality" (2025 edition)

211M changed lines, 2020 to 2024

Moved and refactored lines fell from about 25% of changes to 9.5%. Copy-pasted lines rose from 8.3% to 12.3%, and commits containing five or more duplicated lines rose about 8x during 2024, the headline "4x growth in code clones."

Primary source

GitClear, "The Maintainability Gap"

January 2026, 623M code changes, 2023 to 2026

Code-block duplication +81%, error-masking constructs +47%, cross-file function calls -35%, refactoring's share of all changes fell from 21% in 2022 to 3.8% in 2026. GitClear's own framing: this is deferred structural maintenance, not simply "AI writes bad code."

Primary source

Vero et al., "BaxBench: Can LLMs Generate Correct and Secure Backends?"

February 2025, 392 backend tasks

62% of solutions from the best model at the time were either incorrect or contained a security vulnerability. This counts incorrect-or-insecure, not insecure alone; the 2026 leaderboard sits near 50%.

Primary source

Perry, Srivastava, Kumar, Boneh, "Do Users Write More Insecure Code with AI Assistants?"

CCS 2023, 47 participants

Developers with an AI assistant wrote significantly less secure code and were more likely to believe it was secure. SQL-injectable solutions: 36% with AI vs 7% without.

Primary source

Pearce et al., "Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions"

IEEE S&P 2022

About 40% of 1,689 Copilot-generated programs across 89 CWE scenarios were vulnerable.

Primary source

Veracode, "2025 GenAI Code Security Report"

07-30-2025

45% of samples failed OWASP Top 10 checks (Java 72%); XSS defenses failed 86% of the time and log injection 88%.

Primary source

Veracode, "2026 GenAI Code Security Report"

07-28-2026

44% still fail; the best model passes 68% and most models 50 to 53%. Model size has no effect: XSS pass rate is 15%, log injection 12%. Newer and larger models are not fixing this.

Primary source

Apiiro, "4x Velocity, 10x Vulnerabilities"

September 2025, Fortune 50 telemetry

AI-assisted developers committed 3 to 4x more and generated 10x more security findings; privilege-escalation paths rose 322% and architectural design flaws 153%, while syntax errors fell 76%. The bugs move up the stack, from typos to design.

Primary source

Spracklen et al., "We Have a Package for You!"

USENIX Security 2025, 576,000 samples, 16 models

19.7% of recommended packages did not exist, 205,474 unique hallucinated names. Open-source models hallucinated 21.7% vs about 5.2% for commercial models, and 43% of hallucinated names recurred in every re-query, making them registrable and exploitable.

Primary source

Sonar, "The Coding Personalities of Leading LLMs"

08-13-2025, 4,400+ Java tasks

More than 90% of issues were code smells, and the model with the best pass rate also produced 93% more high-severity bugs than its predecessor. Better benchmarks did not mean safer code.

Primary source

Cloud Security Alliance, "Vibe Coding's Security Debt: The AI-Generated CVE Surge"

04-04-2026

74 CVEs attributed to AI coding tools by Georgia Tech's tracking, with the monthly count rising from 6 in January 2026 to 35 in March.

Primary source

Test Quality

Meta, TestGenEval

ICLR 2025

The leading model at the time reached 35.2% coverage but only an 18.8% mutation score; other frontier models were similar. High coverage did not mean the tests caught real bugs.

Primary source

MUTGEN, mutation-guided LLM test generation

IEEE TSE, 2025-2026

LLM-generated suites reached "100% coverage but only 4% mutation score." Feeding the surviving mutants back to the model as a prompt fixed most of the gap.

Primary source

Meta, "Mutation-Guided LLM-based Test Generation at Meta" (ACH)

01-2025

571 generated hardening tests, 73% accepted by engineers, because each test was proven to kill a specific mutant rather than just running without asserting anything useful.

Primary source

ISSTA 2026 replication on mutation score vs bug detection

2026, 11 models, 101,000 tests

Mutation score correlates strongly with real bug detection (r = 0.86) for regression tests, but only weakly when the code under test is already wrong. Generate tests from the acceptance criteria, not from the implementation.

Primary source

Chu et al., survey of LLM-based unit test generation

11-2025, 115 papers

A 115-paper survey of the LLM-based unit test generation literature, the research base behind the guidance that coverage percentage alone does not prove test quality and mutation scoring is what closes the gap.

Primary source

Standards and Regulation

OWASP Top 10 for LLM Applications

2025 and 2026 editions

The LLM01 to LLM10 threat codes used throughout this library's threat model. A 2026 edition was published in 08-2026; codes cited elsewhere are the 2025 edition unless the 2026 mapping is noted.

Primary source

OWASP Top 10 for Agentic Applications 2026

Published 12-09-2025

The ASI01 to ASI10 threat codes for agent-specific risks: excessive agency, memory and context poisoning, unexpected code execution, and human-agent trust exploitation among them.

Primary source

OWASP "Secure Coding with AI" cheat sheet

OWASP GenAI Security Project

The practical companion to the Top 10 lists, cited as the reference for hardening the specific bug classes AI code gets wrong most often, including XSS, log injection, and error masking.

Primary source

OpenSSF, "Security-Focused Guide for AI Code Assistant Instructions"

08-2025

A ready-made starting block for the security rules a repo instruction file should include: input validation, output encoding, parameterized queries, no secrets in code, and dependency pinning.

Primary source

CISA, "Careful Adoption of Agentic AI Services"

05-2026

The government-issued version of the same MCP security advice: inventory every agent and connector, apply least privilege, log everything, and require human approval for consequential actions.

no public link

MCP specification, security best practices

2026-07-28 revision

Made the protocol core stateless and published best practices covering confused deputy attacks, token passthrough, SSRF via metadata, and localhost redirect impersonation.

Primary source

NIST AI RMF 1.0 and Generative AI Profile

RMF January 2023; AI 600-1, July 2024

A Govern, Map, Measure, Manage framework with genAI-specific risks including confabulation. The RMF itself is under revision with no draft published as of 09-2026.

Primary source

NIST SP 800-218A

July 2024

Secure software development practices for when AI models are part of the development process itself, not just the code they help write.

Primary source

ISO/IEC 42001:2023

Plus ISO/IEC 42005:2025 and 42006:2025

The AI management system standard covering risk assessment, impact assessment, lifecycle controls, transparency, and human oversight.

Primary source

EU AI Act, Regulation (EU) 2024/1689

As amended by the Digital Omnibus, Regulation (EU) 2026/1744

The Digital Omnibus, in force 07-27-2026, moved Annex III high-risk obligations to 12-02-2027 and Annex I product-embedded obligations to 08-02-2028. Coding assistants are not named as high-risk in the Commission's draft classification guidelines.

Primary source

Colorado SB 26-189

Effective 01-01-2027

The re-enacted Colorado AI law, an automated-decision notice and documentation statute, after the original SB 24-205 was stayed by a federal court in April 2026 and then repealed.

Primary source

AICPA TQA 9561

09-10-2026, nonauthoritative

A Technical Q&A addressing a service organization's use of AI in SOC 1 and SOC 2 examinations. Practitioner consensus keeps AI-authored changes on the same CC8.1 change-management path as any other change.

no public link

SLSA v1.2

November 2025

Finalized the Source track. Source L4 requires two-party review, which an agent author plus one human reviewer does not satisfy if the agent is counted as a party.

Primary source

Vendor Documentation Cited

Anthropic, workload identity federation

GA 06-17-2026

A federation issuer for token.actions.githubusercontent.com, a service account, and a federation rule constraining subject, audience, and claims, returning a short-lived, single-use token.

Primary source

Anthropic, Claude Code GitHub Actions

Vendor documentation

The vendor's first-party CI action, the one place a subscription seat's OAuth token is permitted to run; third-party tools must use API keys instead.

Primary source

Anthropic, Claude Code best practices, hooks, memory, sandboxing

Vendor documentation

Documents hooks and permission settings as the enforcement layer for anything that must never happen, since an instruction file like CLAUDE.md is advisory only.

Primary source

Anthropic, "Demystifying evals for AI agents"

01-09-2026

A short primer on grading agent transcripts rather than single outputs, cited as the reference for building an eval fixture set for review and planning agents.

Primary source

Anthropic, "Effective context engineering for AI agents"

09-29-2025

Source of the line-by-line test for an instruction file: would removing this line cause a mistake? If not, cut it.

Primary source

Anthropic, BAA coverage

Vendor documentation

A BAA is available for the API and Enterprise; Claude Code CLI and desktop are covered only with zero data retention, while Claude Code on the web and MCP connectors are not covered.

Primary source

GitHub, Enterprise AI Controls and agent control plane

GA 02-26-2026

Adds an AI Administrator role, per-agent enable and disable policy, an actor_is_agent field in the audit log, an agent session viewer, and MCP server allow and deny lists.

Primary source

GitHub, Copilot code review can approve pull requests

Preview, 09-01-2026, default off

Lets an AI review count toward required approvals. Leaving this on for a Tier 3 path is the exact control failure this library's code-review page warns against.

Primary source

GitHub, cloud agent commit attribution and session logs

03-20-2026

Traces any cloud agent commit to its session log through an Agent-Logs-Url trailer, the reference pattern for agent commit attribution.

Primary source

GitHub, usage-based billing (AI Credits)

04-2026

GitHub's move to AI Credits, part of the industry-wide 2026 shift from per-seat to usage-based billing that this library's cost model page budgets around.

Primary source

GitHub, HIPAA position

Vendor documentation

GitHub does not sign BAAs, and Copilot is not HIPAA-eligible under any plan.

Primary source

OpenAI, HIPAA-eligible products

Vendor documentation

A BAA is available for the API and ChatGPT Enterprise. Codex local is BAA-eligible; Codex cloud is not.

Primary source

Linux Foundation, Agentic AI Foundation

12-09-2025

The foundation stewarding the AGENTS.md convention, now read by 60,000-plus open-source projects.

Primary source

AGENTS.md

Open convention

The cross-tool instruction-file convention for Codex, Cursor, Copilot, Devin, Jules, and most other agents. Claude Code has read it natively since September 2026 when no CLAUDE.md exists.

Primary source

How This Library Is Maintained

This page was verified on September 20, 2026 against the primary sources listed above. The site's scheduled currency sweep re-checks each source on a recurring basis, and this category is reviewed at least monthly for new editions, corrected figures, or retracted claims.

When a cited report is updated or superseded upstream, the number on this page is updated to match, and the date qualifier changes with it. If a figure ever looks stale, the primary source link is the way to check it yourself.

Where It Fits in AIDLC

This library is not a sixth phase - it is the evidence underneath all five. Analyze and Ideate use the productivity and adoption data to set realistic expectations before committing to a rollout. Develop and Launch use the quality, security, and test-quality data to size the review and verification work. Curate is where an organization's own measured baseline gets checked against what this research says is typical.

Frequently Asked Questions

Every source in this library is one the AI SDLC Playbook cites for a specific figure or claim. Selection favored studies with a stated methodology and sample size, such as randomized trials, large telemetry sets, and peer-reviewed papers, over marketing claims, and every number was re-verified against its original publication before being included here.

BaxBench's 62% figure is usually repeated as 'AI code has vulnerabilities' when it actually measures solutions that were incorrect or insecure, not insecure alone. METR's 19% slower finding is often generalized past its scope: 16 experienced maintainers working in their own repositories, an effect METR's own 2026 update called very weak evidence. Read the primary source before repeating either one.

The site's scheduled currency sweep re-checks every source on a recurring basis, and this category is reviewed at least monthly for new editions, corrections, or retracted figures. This library was last verified on September 20, 2026 against the primary sources listed here, and any figure that changes upstream gets corrected on the next pass.

Put the Research to Work

See the productivity paradox worked through in full, or go back to the hub for the ten rules that turn this research into a policy.

The Productivity Paradox Back to the Playbook