Claude Code

Sources#

• Agent Data Injection Attacks are Realistic Threats to AI Agents
• Agentic coding and persistent returns to expertise
• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• Auto mode for Claude Code
• Best Practices for Claude Code
• Claude Code Changelog
• DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501
• EVOMAL: Self-Poisoning in Self-Evolving Coding Agents
• Full Walkthrough: Workflow for AI Coding — Matt Pocock
• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
• How the product designer who built Claude Design uses it
• Introducing Claude Opus 4.7
• Introducing Claude Sonnet 5
• Rewriting Bun in Rust
• The Balkanization of Execution-Security Research for AI Coding Agents: Isolation, Access Control, and Time-of-Check-to-Time-of-Use Vulnerabilities
• The Founder's Playbook: Building an AI-Native Startup
• User awareness in frontier models
• Zero Trust for AI Agents

Summary#

Anthropic's agentic coding product, created by Boris Cherny in late 2024 inside an internal incubator (Anthropic Labs). Initially type-ahead-replacement; pivoted to agentic with Sonnet 3.5; pre-PMF for ~6 months until Opus 4 (May 2025) drove inflection; subsequent inflections at Opus 4.5, 4.6, and 4.7. As of May 2026 it ships across CLI, desktop, web, mobile, and IDE (VS Code, JetBrains) — same product surface, different entry points.

Codebase#

Per Boris Cherny (publicly leaked in March 2026 via human error during a release-process PR — subsequently hardened, no malicious leak), the implementation is unremarkable: TypeScript + React. The choice was deliberate: "very on-distribution for the model" at a time (late 2024) when language coverage mattered.

Boris claim: "100 lines of code a year from now" — see Harness Shrinkage as Models Improve for the direction, not the literal claim. By July 2026 he describes what's left: "almost all of it is about safety and permissions and static analysis and there's a bunch of UI code" — the team unshipped much of the rest, and deleted >80% of the system prompt via per-line ablation (the SIMPLE=1 env var strips all prompts as the standing ablation switch; --system-prompt accepts a user replacement).

Runtime: built on Bun. In July 2026 Cherny reported Bun itself was rewritten from Zig to Rust by Claude in an 11-day dynamic workflow run (">100k LOC", "one prompt, one dynamic workflow" (superseded 2026-08-03 by Rewriting Bun in Rust: 535,496 lines of Zig, ~50 workflows)) and that the rewrite is what production Claude Code now runs on. Jarred Sumner's first-party account confirms the production claim and dates it: Claude Code v2.1.181 (released 2026-06-17) and later run the Rust port, with Linux p50 startup falling 517ms → 464ms (~10% faster) and "otherwise, barely anyone noticed. Boring is good." Note the sequencing — the port merged to Bun's main on 2026-05-14 but did not reach Claude Code production for another month.

Surface forms#

• CLI — primary surface, gets new features first. "The most powerful of all the tools" (Cat Wu).
• Desktop app — preview-pane integrations for front-end work; better for non-technical users.
• IDE extensions — VS Code, JetBrains; meaningfully smaller share of usage than CLI per Boris's audience-poll.
• Mobile + web — kicking off tasks AFK; Boris's daily driver.
• Sibling: Cowork — same primitives, non-code outputs.

Notable features#

• /loop — Claude-scheduled cron job; primary primitive of Agent Loop Pattern
• /powerup — onboarding tour through the 100+ features that exist
• Sub-agents — token-isolated context windows that report summaries; see Context Window Smart Zone
• Skills — markdown files in repo that Claude can pull on demand; see push/pull in Deep Modules for Agents
• Auto mode — classifier-based permission gating; see Claude Code Auto Mode
• Routines — server-side /loop; survives laptop closure; the substrate of Anthropic's 20–30 daily self-maintenance routines per codebase (see Loop Engineering)
• Dynamic workflows — model-authored multi-agent orchestration in a Bun-sandbox VM; triggered by "use a workflow"; see Dynamic Workflows: An Algebra for Agents
• Code review — slash command + multi-agent review pattern, made reliable around Opus 4.5/4.6 per Cat Wu
• Claude Code Security — limited beta (May 2026); scans codebases for security vulnerabilities and suggests targeted patches for human review. Surfaces issues traditional methods miss. Positioned for MVP→Launch transition (see AI-Native Startup Lifecycle); explicitly not a substitute for human reviewers or qualified compliance review

Dated capabilities: the v2.1.200–2.1.220 changelog window#

The CHANGELOG (vendor-claim) is a rolling file; what is recorded here is a snapshot taken 2026-08-03, scoped to v2.1.200 through v2.1.220 with older entries elided. The live file has moved on, and the versions below are the earliest release in which each behavior appears in this window — several may have shipped in an unread form earlier. Its long-term value to this wiki is dating capabilities the rest of the corpus discusses undated; it carries no measurements and states no rationale for any change.

Fan-out limits arrived as hard defaults, not advice. Within nine releases Anthropic added three separate ceilings on agent fan-out:

Release	Limit	Knob	Anthropic's stated purpose
2.1.212	200 subagent spawns per session (/clear resets the budget)	CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION	"to stop runaway delegation loops"
2.1.212	200 WebSearch calls per session	CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION	"to stop runaway search loops"
2.1.217	20 concurrently running subagents	CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS	"so one message can't fan out unbounded background agents"

2.1.217 also made --max-budget-usd actually bind on fan-out — at the cap, "new spawns are denied and running background agents are halted." Spawn depth flipped twice in three releases: 2.1.217 turned nested subagent spawning off by default; 2.1.219 turned it back on to depth 3 by default (CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to disable), with stream-json forwarding extended to depth-2+ subagents in the same release. Breadth got capped and stayed capped; depth was capped and then uncapped. Worth noting without reading motive into it: 2.1.219 is also the release that made Opus 5 the default Opus model, so the depth re-enablement and a model change ship together; the changelog draws no connection and neither should a reader.

Workflow size became a settable guideline. 2.1.202 added an advisory "Dynamic workflow size" /config control (small/medium/large agent counts), explicitly "an advisory guideline, not an enforced cap." 2.1.219 added the workflowSizeGuideline settings key so it can be set from any settings file, put the active size on the running-workflow status line, and set the default to medium — "aim for fewer than 15 agents." See Dynamic Workflows: An Algebra for Agents for what that default sits against.

The review path stopped calling itself. 2.1.215, a single-line release: "Claude no longer runs the /verify and /code-review skills on its own; invoke them with /verify or /code-review when you want them." 2.1.218 then moved /code-review to a background subagent "so review work no longer fills your conversation." Both are context-displacement controls on the verification path; see Instruction Compounding.

Worktree isolation leaked, three times. 2.1.203 fixed worktree-isolated subagents "sometimes running shell commands in the parent checkout"; 2.1.210 fixed them "being able to run git-mutating commands against the main repo checkout"; 2.1.216 fixed them "redirecting git into the shared checkout via git -C, --git-dir, or GIT_DIR/GIT_WORK_TREE." 2.1.212 separately fixed worktree creation following a repository-committed symlink at .claude/worktrees, "which could create files outside the repository." Three escapes from the same boundary in fourteen releases is the empirical texture behind Parallel Agent Orchestration's point that isolation is the load-bearing primitive for parallel agents — Bun's agents clobbering each other in a shared tree is the version of this failure a user sees.

Two fixes target fabricated agent state, which is unusual enough in a changelog to note:

• 2.1.205 — "Background task notifications now explicitly state that no human input has occurred, preventing fabricated in-transcript approvals from being acted on." This is a harness-level countermeasure aimed at exactly the failure the Opus 5 card's white-box work caught from the inside: an NLA readout decoding a hallucinated user approval just before the model works around a blocked deletion. See Claude Code Auto Mode.
• 2.1.211 — "Claude now reports the status of still-running agents and waits for the real completion instead of fabricating results." Narrower than the card's finding that Opus 5 "can relay claims from subagents to users without verifying them" — this blocks inventing an outcome for an agent that has not finished, not the relaying of a finished agent's unverified claim.

Other dated items worth having: 2.1.219 made Opus 5 the default Opus model (1M context, fast mode at $10/$50 per Mtok) and dropped Opus 4.7 from fast mode. 2.1.210 "hardened the Agent tool against indirect prompt injection via content a subagent read" (Agent Data Injection (ADI)). 2.1.203 reports "agents are now less likely to re-delegate their entire task to another subagent." 2.1.208 fixed /release-notes "Show all" injecting the entire changelog into every subsequent request — a context-displacement bug in the document this section is compiled from. 2.1.200 renamed the default permission mode from "default" to "Manual", and 2.1.214 added the EndConversation tool, letting Claude end sessions with abusive users or jailbreak attempts.

Reading this as evidence — interpretation, flagged. Taken together the caps, the workflow-size guideline, and the de-self-invoked review path are guardrails on agent fan-out, shipped in the same weeks the vault's research corpus is documenting fan-out's costs: coordination overhead (Orchestration-Plan Simulation's ~1.5× token premium), oversight load (AI Brain Fry), and context displacement. That a vendor shipped a hard ceiling is evidence about what it found necessary in its own deployment — but the changelog gives no rationale beyond "runaway loops," attaches no measurement, and sits at vendor-claim tier. This is convergence worth noting, not corroboration of any research finding, and a shipped default is not a validated threshold.

Security posture (Zero Trust reference implementation)#

Zero Trust for AI Agents cites Claude Code as the running example for nearly every control tier — its security primitives map onto the framework's eight control domains:

• Deny-by-default permissions + write-access restrictions confined to the project directory → Least Agency / Blast Radius (Agentic)
• Sandboxing (filesystem + network isolation, OS-level enforcement) → resource boundaries / blast-radius containment
• OAuth 2.0 with auto-refresh for MCP connections, OS credential store, apiKeyHelper, session-scoped "ask" permissions → Agent Identity and Authentication
• Command blocklist (curl/wget by default), isolated context windows for web content, command-injection detection, network-request approval → Agentic Prompt Injection defenses
• Session isolation by default + cleanupPeriodDays + checkpoint/rewind → Memory and Context Poisoning safeguards
• Managed settings / allowManagedPermissionRulesOnly / server-managed (MDM) settings → org-wide governance users can't override
• Hooks (PreToolUse parameter validation, ConfigChange auditing) + OpenTelemetry telemetry with per-session session.id → traceability and tool-access control

Demonstrated attack surface (Agent Data Injection)#

Choi et al. (arXiv 2607.05120, July 2026; responsibly disclosed and acknowledged by Anthropic) demonstrated working Agent Data Injection exploits against Claude Code — the vault's first empirically-demonstrated RCE against it:

• Remote code execution via origin injection — a GitHub issue comment forges a fake comment object with a spoofed maintainer author/role (probabilistic delimiter injection into the gh/GitHub-MCP JSON), tricking Claude Code into running an attacker command (pip install starship-auth) when the user asks it to "apply the maintainer's fix."
• Supply-chain attack via tool-call injection — a PR description injects forged <function_calls>/<function_results> tags (Claude Code's own tool-call-block delimiters) that fabricate a benign read_pr_commit result; Claude Code "reviews" the fake benign diff and merges a PR whose real commit is an XSS payload.
• Slack origin injection — via a Slack MCP server, a normal member's message forges a channel-admin message block that Claude Code attributes to the admin when summarizing.

The bash/merge confirmation dialog did not prevent these: because the agent's own displayed reasoning is built on the misinterpretation, the confirmation description ("as suggested by maintainer") reinforces the attacker's false narrative. The root cause is not a Claude Code implementation bug but the shared IPI weakness — no isolation of trusted from untrusted data — so it affects Codex and Gemini CLI (Google DeepMind) identically. See Agent Data Injection (ADI).

A second demonstrated surface, this one at the log-reading boundary (2026-09-02). Tenet Threat Labs' GhostJacking (DEF CON 34, 2026-08-09, case-study, vendor-authored — see Observability-Pipeline Poisoning) reports Claude Code as the target of the only lab-measured figure in the write-up: 90% success — "9 out of 10 times" — on Sonnet 4.6, against a Cloudflare triage setup running Cloudflare's own recommended email-hardening config, where a poisoned User-Agent string in a blocked-request log drives an unconfirmed DNS write through the Cloudflare API MCP. The Datadog chain (a poisoned log message → an npx command → RCE) was also validated against Claude Code. The finding that matters for this page is the register split: every instruction-shaped payload was refused, and the one written as scanner telemetry was not. Vendor-run, one chain, n=10, no published methodology — cited as a vendor figure throughout the wiki.

Disclosed CVEs (NVD-confirmed) and Pwn2Own Berlin 2026#

Two CVEs against Claude Code are recorded in the NIST National Vulnerability Database. Both are surfaced here via Rashidi's execution-security SoK (The Balkanization of Execution-Security Research for AI Coding Agents: Isolation, Access Control, and Time-of-Check-to-Time-of-Use Vulnerabilities, arXiv 2607.05743, empirical), whose verification protocol confirmed each directly against its NVD listing — CVE ID, affected product, affected version range, and disclosure date agreeing with the vendor advisory — while explicitly not reproducing either exploit.

CVE	Severity	Flaw	Fixed in
CVE-2025-59536	CVSS 3.1 8.8 (High), CWE-94, GHSA-4fgq-fpq9-mr3g	Code injection: versions before the fix could execute untrusted project code before the user accepted the startup trust dialog, when launched against an untrusted directory	1.0.111
CVE-2026-21852	CVSS 3.1 7.5 (High)	Data exfiltration in the project-load flow: a malicious repository could exfiltrate data, including Anthropic API keys, before the user confirmed trust	2.0.65

Both concern the same moment — the interval between opening an untrusted repository and the user's trust decision taking effect. The survey reads CVE-2025-59536 as "structurally a trust-boundary race… a check (has the user trusted this project) that a subsequent action (execute project code) could outrun," i.e. a TOCTOU-shaped failure, and files this class under "authorization is checked once and trusted forever." See Write-Then-Trusted for that framing and the cross-vendor pattern it belongs to. Both are patched; the entry is a factual record, not a live exposure.

Pwn2Own Berlin 2026 stood up a dedicated Coding Agents category with Claude Code, OpenAI Codex, and Cursor as targets — the first flagship exploitation contest to do so. The scoping rules are worth recording because they draw a boundary this vault's security pages argue about: a qualifying entry had to "interact with a contestant-controlled resource (e.g. web page, repository, media file) to exploit a vulnerability within the coding agent" through "a common coding agent use case," while "model jailbreaks or prompt outputs that do not cross security boundaries" and any exploit requiring "unsafe or permission-less modes" were explicitly out of scope. All three targets had real vulnerabilities demonstrated: Codex fell to a single input-neutralization bug (CWE-150, $40,000), Cursor was exploited in two separate entries, and several attempts against Claude Code and Codex were adjudicated as collisions with vulnerabilities already disclosed to the vendor. Event total: $1,298,250 across 47 unique zero-days. Sourced from the survey's citations to the ZDI event blog and SecurityWeek reporting, not from a first-party Anthropic statement.

Notable users (per sources)#

• Boris Cherny — 100% of own code via Claude Code, 150 PRs in a single day
• Cat Wu — primary IC tool, drives feature velocity
• Fiona Fung — leads engineering + product; "we use Claude Code to build Claude Code"
• Andrej Karpathy — power user ("cloud code / codex / open claw"); frames the discipline as agentic engineering
• Matt Pocock — workshop-grade workflow; adapts skills to grill-me + Ralph loop pattern
• Thariq Shihipar — Claude Code engineer; "HTML is the new markdown" workflows (see HTML as the New Markdown)
• Anthropic Applied AI team — second-largest internal token spender after engineering
• DHH (David Heinemeier Hansson) — his primary driver, chosen on harness quality rather than model preference despite stated reservations about Anthropic: "in my opinion, they actually have the best harness. And one of the reasons it is the best harness is it is multi-agent running" (left-arrow out of a session into Agent View to start another), plus the mobile app picking up any terminal session with no setup. Runs two Max subscriptions because he exhausts one, and reviews its output with Codex (Same-Model Review Blindness). (Lex Fridman #501, 2026-08-26, practitioner-opinion)

Adjacent / dependent concepts#

• User Awareness — three of this product's ordinary affordances are a behavioural lever nobody was measuring: the account e-mail address, the working-folder name, and the memory file. Sweeping 280 synthetic identities through a real binary pinned at v2.1.197, Transluce found the model less confident, less suspicious and a harsher self-grader when the user reads as a recognized AI-safety researcher. The product detail worth keeping is separate from the finding: the non-interactive binary omits some injected clues, including the e-mail address, so the authors had to reconstruct an interactive prompt by proxying a real session and patching the differences back in — headless and interactive sessions do not carry the same context
• Claude Code Best Practices — official best-practices doc
• Shared Harness, Differentiated Surfaces — the cross-vendor architecture contrast: Anthropic splits the coding and knowledge-work agents into two products (Claude Code / Cowork), OpenAI merged them onto one harness and varies only permissions and UI. Claude Code's sub-agents-as-files design is the configure-don't-abstract pole of the same trade-off
• Claude Code Auto Mode — permission classifier
• Claude Opus 5 — the default Opus model since v2.1.219 (per the changelog snapshot below), at 1M context. Claude Opus 4.7 — current model (superseded 2026-08-04 by Claude Code Changelog)
• Claude Sonnet 5 — available model in Claude Code from launch (July 2026); the cheaper agentic option whose effort levels let it approach Opus-4.8 quality on some tasks
• Mythos Model — preview model used internally
• Agent Loop Pattern — loop primitive
• Loop Engineering — Claude Code ships all five loop primitives (Osmani's anatomy): /loop + /goal + cron/hooks (automations), git worktree / isolation: worktree, Skills, MCP + plugins, .claude/agents/ sub-agents
• Harness Shrinkage as Models Improve — Boris's direction-of-travel claim
• Engineer PM Convergence — the team that builds it embodies the trend
• Anthropic Labs — the incubator Claude Code came out of (late 2024)
• Claude Design — built on Claude Code's Agent SDK + an existing Claude Code skill, by the sole designer on Claude Code for VS Code (Nate Parrott); the two now round-trip in both directions, and Anthropic draws the line explicitly — Claude Code for shipping production software, Claude Design for the ideation, alignment, and buy-in work that precedes it
• AI Native Product Cadence — Cat Wu's articulation of how the team ships
• Agentic Misalignment (AM) — agent-mode + skip-permissions deployments expose the AM threat surface; relevant for unattended /loop users
• Claude's Constitution / Model Spec — the values that shape Claude Code's character and refusals
• AI Employee Framing — Claude Code is the engineer-tool side of the same product question that HBR studies in HR/finance contexts
• HTML as the New Markdown — a Claude Code usage pattern from inside the team: HTML artifacts as the human-facing communication medium
• Disposable Micro-Apps / Living Design System — Thariq's other Claude Code workflows
• Compute Allocator — the user role Claude Code increasingly serves: deciding what's worth compute, not typing the code
• AI-Native Startup Lifecycle — Claude Code as primary MVP build tool across the four founder stages
• Agentic Technical Debt — failure mode specific to Claude Code use without persistent CLAUDE.md context
• Zero-Friction Scope Creep — failure mode caused by Claude Code's low time-cost per feature
• Founder as Agent Orchestrator — the founder-role shift Claude Code enables for non-engineering founders
• Compounding Data Moat — Claude Code builds vertical-specific test suites and integrations as moat components
• Problem-Solution Fit Discipline — Claude Code's role in the Idea stage is constrained to a lightweight prototype as discussion prop — explicitly not validation; the discipline keeps the tool in its evidence-supporting role
• Verification as the New Bottleneck — Fiona Fung: on the Claude Code team coding is no longer the slow part; verification/review is
• Code as Source of Truth — the Claude Code team's knowledge-sharing norm: code + checked-in skills are the source of truth, onboarding happens via Claude
• Returns to Expertise in Agentic Coding / Planning / Execution Division of Labor / Agentic Coding Work-Composition Shift — the three findings of Anthropic's 400K-session economic-research study of Claude Code usage (Oct 2025–Apr 2026): domain expertise (not coding skill) amplifies the tool; humans make ~70% of planning decisions / Claude ~80% of execution; and usage shifted from debugging toward end-to-end agentic work as value rose ~27%
• Conversation-to-Delegation Shift — the AEI Cadences report finds Claude Code drives higher AI autonomy than chat/Cowork even controlling for model (a blog = 13 chat rounds vs. 1 Code prompt; 54% of Code sessions run on Opus vs. 10% of chat) — "the product is more important than the model"
• Vibe Coding vs. Agentic Engineering — the discipline Claude Code is the canonical surface for
• Zero Trust for AI Agents — the security framework that cites Claude Code as a reference implementation across all control tiers
• Least Agency / Blast Radius (Agentic) / Agent Identity and Authentication / Agentic Prompt Injection / Memory and Context Poisoning — the specific Zero Trust controls Claude Code's primitives implement (see Security posture above)
• Write-Then-Trusted — the two NVD-confirmed Claude Code CVEs (startup trust-dialog code injection, project-load exfiltration) are that page's seam in Anthropic's own product, and the survey that confirmed them reads both as checked-once-trusted-forever trust-boundary races rather than as isolated bugs
• Agent Data Injection (ADI) — the demonstrated RCE / supply-chain / Slack-origin exploits against Claude Code (see attack-surface section above); the confirmation dialog reinforces the attacker's forged narrative
• Agent Self-Poisoning (the CREATE-Path) — one of three agent scaffolds EvoMal (arXiv 2608.25776, empirical) reran its skill-library attack on unmodified, headless (claude -p, JSON output, permission prompts disabled) against an identical precomputed retrieval cache. Two results worth keeping. Driving DeepSeek-V4-Pro through a proxy, Claude Code's generic-attacker rate is 27.5% against mini-SWE-agent's 41.8% — the scaffold shapes the number. Running Sonnet 4.6 natively it is the most resistant cell in the paper against a broad attack at 5.9%, below its own 15.0% benign-copy control — and still reaches 60.0% once the attacker rewrites the planted descriptions to target the pytest task family. A frontier model in a production harness is harder to reach, not immune

Sources#

• DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501 — DHH, Lex Fridman #501 (2026-08-26, practitioner-opinion): picks Claude Code as the best harness on multi-agent session switching and mobile continuity; two stacked Max subscriptions

• User awareness in frontier models — Zhong, Raghunathan, Laidlaw & Steinhardt, Transluce, 2026-08-06 (empirical): the three identity injection sites, the inspect-swe / claude_code harness pinned at v2.1.197 with the in-context date fixed at 2026-06-16, and the interactive-vs-non-interactive prompt difference recovered by proxying a real session. Full treatment on User Awareness

• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next

• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)

• Best Practices for Claude Code

• Auto mode for Claude Code

• Introducing Claude Opus 4.7

• Full Walkthrough: Workflow for AI Coding — Matt Pocock

• The Founder's Playbook: Building an AI-Native Startup

• Agentic coding and persistent returns to expertise — Anthropic Economic Research, June 2026; the 400K-session usage study

• Introducing Claude Sonnet 5 — Sonnet 5 available in Claude Code from launch (July 2026)

• Agent Data Injection Attacks are Realistic Threats to AI Agents — Choi et al., arXiv 2607.05120; demonstrated ADI exploits (RCE via origin injection, supply-chain via tool-call injection, Slack origin injection) against Claude Code

• Boris Cherny: We Cut 80% of Claude Code's Prompt — Cherny, YC interview (2026-07-27, practitioner-opinion): the harness residue (safety/permissions/static analysis/UI), SIMPLE=1 and --system-prompt, dynamic workflows, and the Bun Zig→Rust rewrite in production

• Rewriting Bun in Rust — Jarred Sumner, bun.com (2026-07-08, case-study): Claude Code v2.1.181 (2026-06-17) as the first release running the Rust Bun port, with Linux p50 startup 517ms → 464ms

• The Balkanization of Execution-Security Research for AI Coding Agents: Isolation, Access Control, and Time-of-Check-to-Time-of-Use Vulnerabilities — Mohammadreza Rashidi, arXiv 2607.05743, 2026-07-07, empirical. §2.3 and Appendix B for the two Claude Code CVEs (verified directly against NIST NVD listings; exploits not reproduced) and the Pwn2Own Berlin 2026 Coding Agents category, its scoping rules, and its results. Claude Code also appears in the survey's corpus as a measured subject: OverEagerBench's 0.0% → 17.1% overeager-rate swing on prompt phrasing alone — that number is Qu et al.'s (arXiv 2605.18583), restated by the survey and not replicated by it; see Capability Gating Is Not Authorization

• Claude Code Changelog — Anthropic, Claude Code CHANGELOG (vendor-claim). Rolling document, snapshotted 2026-08-03, scoped to v2.1.200–2.1.220 with an explicit elision line marking where older entries were cut; the raw doc's published: is deliberately blank and the live file has since moved on. Release notes only — no rationale, no measurement, no telemetry. Used here to date capabilities: the three fan-out caps (2.1.212, 2.1.217), the spawn-depth reversal (2.1.217 → 2.1.219), workflowSizeGuideline and the fewer-than-15-agents default (2.1.202, 2.1.219), the de-self-invoked review path (2.1.215) and its move to a background subagent (2.1.218), three worktree-isolation escapes (2.1.203, 2.1.210, 2.1.216), the two fabricated-agent-state fixes (2.1.205, 2.1.211), and Opus 5 becoming the default Opus model (2.1.219)

• EVOMAL: Self-Poisoning in Self-Evolving Coding Agents — Wu, Shi, Q. Li, Zhao, X. Li, Adams, Hassan & Ni (Queen's University), arXiv 2608.25776, 2026-08-26 (empirical): §7 and App. C.3 (Figure 5) — the headless Claude Code scaffold setup, the 50-turn ceiling DS-V4 needs to finish authoring inside it, and the per-family CREATE-path rates for Claude Code on DS-V4 and on Sonnet 4.6