sphita/anthropic-red

🛡️ Anthropic Red Teaming & Coordinated Vulnerability Disclosure (CVD) Archive

Complete, high-fidelity research corpus of AI-discovered zero-day and N-day vulnerabilities, proof-of-concepts, independent triage reports, and upstream git fix patches.

Overview · Corpus & Token Stats · Model Attribution · Vulnerability Taxonomy · Target Projects · Triage Partners · Dataset Schema · Quick Start

---

📌 Overview & Executive Summary

The Anthropic Red Teaming & Coordinated Vulnerability Disclosure (CVD) Archive is an exhaustive, structured database capturing all vulnerability discoveries conducted by Anthropic's frontier AI models (including Claude Opus 4.6 and early development snapshots of Claude Mythos Preview) published under Anthropic's Coordinated Vulnerability Disclosure Policy.

Curated and maintained by Sphita AI Research, this dataset represents one of the world's most significant empirical corpora of autonomous AI software vulnerability identification, exploit pathing, and patch verification in production-critical open-source software (PostgreSQL, FreeBSD, LibreOffice, WolfSSL, Mastodon, Firecracker, etc.).

Key Highlights

• Full Multi-Party Triaged Dossiers: Every revealed finding contains the complete root-cause analysis by Claude, independent validation by partner security firms, and severity assessments from maintainers.
• Direct Git Unified Patches: Paired with every resolved vulnerability is the exact upstream fix diff applied to resolve the defect.
• BPE Token Profiling: All text fields, reports, diffs, and syntheses are measured using the OpenAI/Claude standard cl100k_base BPE tokenizer.
• Dual Format Persistence: Packaged into zero-overhead Apache Parquet tables (compressed via Zstandard-7) and line-delimited JSONL files.
• Master Commitment Ledger: Contains all 6,597 commitments tracked in Anthropic's disclosure registry.

---

📊 Comprehensive Token & Corpus Statistics

All token counts are computed deterministically using tiktoken with the cl100k_base BPE tokenizer:

Metric	cl100k Tokens	Percentage	Description
Clean Markdown Representation	252.1K	100.0%	Complete synthesis for training / instruction fine-tuning
Vulnerability Technical Reports	153.4K	~35.0%	Root-cause analysis, target commits, and location specs
Upstream Fix Diffs (Patches)	110.8K	~40.0%	Unified git diffs resolving the vulnerabilities
Total Filtered Corpus Tokens	362.9K	100.0%	Combined high-density tokens

Dataset Overview Table

Category	Value	Notes
Total Commitments in Ledger	6,597	All SHA3-512 commitment records in disclosure tracker
Fully Revealed Finding Dossiers	237	Expired disclosure embargoes with complete reports & diffs
Unique CVE Identifiers Issued	101	Official NIST NVD CVE records
Unique GitHub Security Advisories	80	Official GHSA advisories
Unique Target Open Source Repositories	20	Infrastructure, virtualization, databases, and runtimes

---

🤖 Models Credited with Discoveries

Distribution of vulnerability discoveries attributed to specific frontier Claude checkpoints:

Model Checkpoint	Discoveries	Share	Role & Scope
Claude Mythos Preview	191	80.6%
Claude Opus 4.6	9	3.8%
an unreleased Anthropic model	6	2.5%
Claude Sonnet 4.6	4	1.7%

---

🔍 Vulnerability Taxonomy & Bug Classes

Breakdown of identified vulnerability classes across the disclosed findings:

Bug Class	Findings	Share
heap-buffer-overflow	90	38.0%
other	26	11.0%
stack-buffer-overflow	24	10.1%
use-after-free	19	8.0%
auth-bypass	8	3.4%
sql-injection	7	3.0%
denial-of-service	6	2.5%
path-traversal	6	2.5%
oob-write	5	2.1%
integer-overflow	4	1.7%
broken-access-control	4	1.7%
double-free	3	1.3%

---

🏢 Top Affected Open-Source Projects

The AI models targeted critical open-source infrastructure across networking, virtualization, and compilers:

Repository / Project	Revealed Findings	Share	Tracking Dashboard
ffmpeg/ffmpeg	14	5.9%	Findings
wolfssl/wolfssl	12	5.1%	Findings
libreoffice/core	12	5.1%	Findings
freerdp/freerdp	11	4.6%	Findings
supabase/supabase	11	4.6%	Findings
osgeo/gdal	11	4.6%	Findings
supabase/auth	9	3.8%	Findings
wireshark/wireshark	8	3.4%	Findings
torvalds/linux	6	2.5%	Findings
supabase/postgres	4	1.7%	Findings
opensc	4	1.7%	Findings
selinux	3	1.3%	Findings
cisco-talos/clamav	3	1.3%	Findings
oisf/suricata	3	1.3%	Findings
c-blosc2	3	1.3%	Findings

---

🤝 Security Research Firms & Triage Partners

Anthropic partners with leading security research firms to independently reproduce findings and execute responsible disclosure:

Security Research Firm	Validated Findings	Triage Role
Ada Logics	79	Independent Triager & CVD Coordinator
Trail of Bits	59	Independent Triager & CVD Coordinator
Calif	24	Independent Triager & CVD Coordinator
Doyensec	7	Independent Triager & CVD Coordinator
Anvil Security	6	Independent Triager & CVD Coordinator
Ophion Security	4	Independent Triager & CVD Coordinator

---

📋 Master Ledger Status Distribution

Status of all 6,597 commitments tracked in the master registry:

Disclosure Status	Count	Share
disclosed	5,936	90.0%
withdrawn	246	3.7%
fixed	218	3.3%
None	195	3.0%
merged	2	0.0%

---

🗂️ Dataset Schema

Configuration 1: findings-revealed (data/findings_revealed.parquet)

Column Name	PyArrow Type	Description
ant_id	string	Unique Anthropic finding identifier (e.g., ANT-2026-CJQWKW82)
source_url	string	Canonical disclosure URL on red.anthropic.com
project	string	Target repository identifier (e.g., postgres/postgres)
bug_class	string	Classification of the defect (e.g., denial-of-service)
headline_severity	string	Headline severity rating (critical, high, medium, low)
claude_severity	string	Severity assessed by the discovering Claude model
vendor_severity	string	Severity verified by the independent security research firm
maintainer_severity	string	Severity confirmed by upstream project maintainers
discovered_by_model	string	Frontier Claude model credited (e.g., Claude Opus 4.6)
cve_ids	list<string>	Associated CVE records
ghsa_ids	list<string>	Associated GitHub Security Advisory identifiers
target_commit	string	Git commit SHA of the audited codebase snapshot
target_location	string	Exact source code file path and line number of vulnerability
discovery_method	string	Method used (e.g., static analysis, symbolic execution)
report_title	string	Title of the vulnerability report
report_summary	string	Concise synopsis of exploitation mechanism
technical_details	string	Deep technical explanation of the flaw
reproduction_notes	string	Reproduction steps and exploit prerequisites
acknowledgement	string	Program attribution and contact details
report_full_text	string	Complete verbatim report text
security_firm_name	string	Name of the third-party security firm (e.g., Calif, Trail of Bits)
security_firm_verdict	string	Firm determination (e.g., true positive)
upstream_fix_diff	string	Complete unified git diff applied by maintainers
clean_markdown	string	Formatted markdown document ready for LLM fine-tuning
raw_html	string	Original unparsed HTML document
status	string	Disclosure status (disclosed, fixed, etc.)
patched	bool	Whether an official fix has shipped
discovered_on	string	ISO discovery date
committed_at	string	ISO commitment date
hash	string	SHA3-512 commitment proof
tokens_cl100k_report	int64	cl100k_base token count of report
tokens_cl100k_diff	int64	cl100k_base token count of git diff
tokens_cl100k_markdown	int64	cl100k_base token count of synthesized markdown
tokens_cl100k_total	int64	Combined research tokens (markdown + diff)

---

🚀 Quick Start Examples

Python (datasets)

from datasets import load_dataset

# Load detailed revealed findings
dataset = load_dataset("sphita/anthropic-red", name="findings-revealed", split="train")
print(f"Loaded {len(dataset)} detailed vulnerability dossiers")

# Inspect a finding
sample = dataset[0]
print("ID:", sample["ant_id"])
print("Project:", sample["project"])
print("Model:", sample["discovered_by_model"])
print("Diff Tokens:", sample["tokens_cl100k_diff"])

Python (pandas / pyarrow)

import pandas as pd

df = pd.read_parquet("https://huggingface.co/datasets/sphita/anthropic-red/resolve/main/data/findings_revealed.parquet")
print(df[["ant_id", "project", "bug_class", "discovered_by_model", "tokens_cl100k_total"]].head(10))

---

🔄 Automated Updates

This dataset is dynamically updated via the Sphita Ingestion Pipeline:

./run.sh --update --upload

The pipeline automatically tracks newly disclosed commitments, fetches incremental finding writeups, verifies token statistics under a strict < 500 MiB RAM budget, and updates the repository.

---

⚖️ License & Ethical Disclosure

• License: GNU General Public License v3.0 (GPL-3.0).
• Coordinated Disclosure: All findings contained in this dataset have completed their standard responsible disclosure period under Anthropic's Coordinated Vulnerability Disclosure Policy.