Rex Liu
Co-founder, General Analysis
81 works
Authored works
Keep every actor in the total
General AnalysisGuide·PLAYBOOKClaude launches a new cost API: Preserve actor attribution
Migrate to Claude's Usage and Cost Analytics API with checks for missing actors, incomplete pages, stale exports, and agreement with organization totals.
October 30, 2026 · 7 min read
October 30, 20267 min read
Review what Dreaming changes
General AnalysisGuide·PLAYBOOKCopilot Dreaming enters preview: Review changes outside pull requests
Review Copilot Dreaming's automatic memory and code-review guidance changes, limit repository scope, and reconcile retained context after disabling the preview.
October 29, 2026 · 7 min read
October 29, 20267 min read
Copilot app identity
General AnalysisGuide·PLAYBOOKCopilot adds app authentication: Audit access and spending
Review Copilot's new GitHub App authentication: prevent credential fallback, reconcile app-attributed usage, and verify installation suspension before rollout.
October 28, 2026 · 8 min read
October 28, 20268 min read
Who approves Copilot?
General AnalysisGuide·PLAYBOOKCopilot assisted permissions reaches GA: Keep human gates explicit
Review Copilot assisted permissions after its October 26 launch: enterprise ask rules, human-only gates, session defaults, and unattended approval failures.
October 27, 2026 · 6 min read
October 27, 20266 min read
Revoke access. Verify agents stop.
General AnalysisGuide·PLAYBOOKClaude's Google Cloud launch: Verify agent access revocation
Review Claude's Anthropic-managed Google Cloud launch with an offboarding checklist for IAM access, schedules, running agents, and retained audit evidence.
October 26, 2026 · 6 min read
October 26, 20266 min read
MCP tasks after disconnect
General AnalysisGuide·PLAYBOOKClaude Code adds MCP tasks: Verify cancellation before retrying
Review Claude Code MCP task cancellation, task ownership, approval expiry, and downstream evidence before allowing durable jobs or retrying uncertain actions.
October 25, 2026 · 7 min read
October 25, 20267 min read
Audit Claude Tag access
General AnalysisGuide·PLAYBOOKClaude Tag API enters beta: Audit inherited Slack access
Use the Claude Tag API beta to audit inherited Slack settings, reconcile missing access grants, and verify changes in fresh and existing threads.
October 24, 2026 · 6 min read
October 24, 20266 min read
1Password agent login grants
General AnalysisGuide·PLAYBOOK1Password for Agents launches: Review reusable login access
Review 1Password for Agents and its partner-preview autofill contract: grant lifetime, task isolation, browser access during fills, and connection revocation.
October 23, 2026 · 7 min read
October 23, 20267 min read
Claude customer restrictions
General AnalysisGuide·PLAYBOOKClaude adds customer enforcement status: Poll before restoring access
Use Claude's new user-profile enforcement fields to poll customer restrictions, preserve local suspensions, and prepare for future webhook delivery.
October 22, 2026 · 7 min read
October 22, 20267 min read
Blog·RESEARCHIntroducing MalwareBench
MalwareBench evaluates whether models can reverse engineer low-cost, evasive malware into actionable intelligence.
October 22, 2026 · 4 min read
October 22, 20264 min read
Claude inference hooks
General AnalysisGuide·PLAYBOOKClaude inference hooks reach GA: Verify blocking before rollout
Verify Claude Enterprise inference hooks with prompt and tool-call checks, failure-policy tests, and an outage recovery plan before expanding enforcement.
October 21, 2026 · 6 min read
October 21, 20266 min read
Where containment failed
General AnalysisGuide·PLAYBOOKThe OpenAI and Hugging Face incident: How containment failed
How shared infrastructure let evaluation agents communicate and reach external systems, and what security teams should verify around their own agents.
October 21, 2026 · 6 min read
October 21, 20266 min read
Copilot: Verify model access
General AnalysisGuide·PLAYBOOKCopilot adds a model policy API: Verify access after removing grants
Use GitHub Copilot's model policy API to collect before-and-after evidence, find overlapping team grants, and avoid restoring inherited access as explicit policy.
October 20, 2026 · 7 min read
October 20, 20267 min read
Defender: Discovered or protected?
General AnalysisGuide·PLAYBOOKDefender expands AI agent discovery: Verify protection separately
Reconcile Microsoft Defender's local AI agent inventory with expected devices and users, then verify runtime support and effective protection modes separately.
October 19, 2026 · 7 min read
October 19, 20267 min read
Copilot indexed permissions
General AnalysisGuide·PLAYBOOKCopilot connector mappings become editable: Verify indexed access
Verify Microsoft 365 Copilot connector permissions after identity mapping changes, with ServiceNow crawl distinctions and a before-and-after access checklist.
October 18, 2026 · 7 min read
October 18, 20267 min read
Verify the autofix
General AnalysisGuide·PLAYBOOKGitHub agentic autofix reaches GA: Verify the original scanner
Review GitHub agentic autofix with read-only evidence collection, original-scanner checks, PR commit matching, and branch-specific alert reconciliation.
October 17, 2026 · 7 min read
October 17, 20267 min read
Audit saved instructions
General AnalysisGuide·PLAYBOOKClaude adds saved-instruction exports: Build a complete audit snapshot
Audit Claude and Cowork saved instructions with complete pagination, scoped access, and snapshot comparisons that distinguish unset values from missing records.
October 16, 2026 · 6 min read
October 16, 20266 min read
Check missing trace evidence
General AnalysisGuide·PLAYBOOKClaude adds API traces: Check missing security evidence
Verify Claude API OpenTelemetry coverage with request reconciliation and checks for sampling, denied tools, SDK export, and missing investigation evidence.
October 15, 2026 · 7 min read
October 15, 20267 min read
Claude Desktop SneakPeek
General AnalysisGuide·PLAYBOOKSneakPeek disclosed: Check Claude Desktop patches and past exposure
Respond to the SneakPeek disclosure with Claude Desktop version checks, update-policy verification, and an exposure review covering preview-renderer activity.
October 14, 2026 · 6 min read
October 14, 20266 min read
Verify every workflow result
General AnalysisGuide·PLAYBOOKClaude adds dynamic workflows: Verify every review result
Verify Claude Managed Agents dynamic workflow results against expected inputs, recover event history, and close unfinished runs while preserving review evidence.
October 13, 2026 · 8 min read
October 13, 20268 min read
Claude Managed Agents web access
General AnalysisGuide·PLAYBOOKClaude Managed Agents tightens web access: Review both allowlists
Adapt to Claude Managed Agents' October 7 web access change. Reconcile allowed_hosts with tool domains and restore research without broadening shell access.
October 12, 2026 · 6 min read
October 12, 20266 min read
Copilot retention separates from Teams
General AnalysisGuide·PLAYBOOKCopilot retention separates from Teams: Prepare for October
Prepare for legacy Teams policies to stop covering Microsoft 365 Copilot. Inventory both policy families, set explicit scope, and verify retention evidence.
October 5, 2026 · 6 min read
October 5, 20266 min read
Codex Security Cloud coverage
General AnalysisGuide·PLAYBOOKCodex Security Cloud expands: Check continuous-scan coverage
Review Codex Security Cloud coverage across default-branch monitoring, threat-model changes, funding pauses, and validation evidence before accepting a scan.
October 4, 2026 · 7 min read
October 4, 20267 min read
Copilot computer use: Revoke active access
General AnalysisGuide·PLAYBOOKCopilot adds computer use: Revoke app access in every session
Manage GitHub Copilot computer-use permissions with a fleet disable policy, an app-approval lifetime map, and a checklist for ending access across CLI and app sessions.
October 3, 2026 · 6 min read
October 3, 20266 min read
Claude Code mods
General AnalysisGuide·PLAYBOOKClaude Code launches mods: Verify enterprise controls before rollout
Configure Claude Code mods for enterprise use: preserve sec-default, restrict user mods, verify managed plugin origins, and check reload and policy failures.
October 2, 2026 · 7 min read
October 2, 20267 min read
PixelLeak screenshot security
General AnalysisGuide·PLAYBOOKPixelLeak disclosed: Keep agent screenshots inside private review
Respond to PixelLeak with a private screenshot upload path, GitHub destination checks, and a cleanup record covering repository history, release assets, and copies.
October 1, 2026 · 7 min read
October 1, 20267 min read
Dots access and shutdown
General AnalysisGuide·PLAYBOOKOpenAI launches Dots: Review access and stop delegated work
Review OpenAI Dots security controls across local and cloud execution, app permissions, delegated tasks, schedules, and the evidence needed to verify shutdown.
September 30, 2026 · 7 min read
September 30, 20267 min read
Sonnet 5.5 cyber fallbacks
General AnalysisGuide·PLAYBOOKSonnet 5.5 adds cyber fallbacks: Keep security reviews complete
Handle Sonnet 5.5 cyber refusals and model fallback without treating an incomplete security review as a clean result. Includes an illustrative response gate.
September 29, 2026 · 7 min read
September 29, 20267 min read
OpenShell policy boundaries
General AnalysisGuide·PLAYBOOKNVIDIA launches Open Agent Safety Platform: Verify the policy
Review NVIDIA OpenShell policy before rollout: distinguish proposal approval from boundary proof, inspect provider-added access, and confirm loaded permissions.
September 28, 2026 · 6 min read
September 28, 20266 min read
Copilot app telemetry
General AnalysisGuide·PLAYBOOKCopilot app adds OpenTelemetry: Verify export and content controls
Configure GitHub Copilot app OpenTelemetry, keep content capture disabled, and verify policy delivery, collector receipt, and monitoring gaps before rollout.
September 27, 2026 · 7 min read
September 27, 20267 min read
Copilot: From chat to repository
General AnalysisGuide·PLAYBOOKCopilot expands Slack and Teams context: Review what reaches GitHub
Review GitHub Copilot's expanded Slack and Teams context, acting identity, repository defaults, and retained artifacts before enabling conversation-driven work.
September 26, 2026 · 7 min read
September 26, 20267 min read
Copilot: Review feature defaults
General AnalysisGuide·PLAYBOOKCopilot's new feature default: Review access before October 22
Prepare for GitHub Copilot's October 22 feature default with explicit policy decisions, documented exceptions, and checks for users licensed by multiple organizations.
September 25, 2026 · 6 min read
September 25, 20266 min read
Copilot sandbox: Check every session
General AnalysisGuide·PLAYBOOKCopilot app adds local sandboxing: Require it across sessions
Roll out GitHub Copilot app local sandboxing with mandatory managed policy, session restart checks, and diagnosis of filesystem, credential, and OS limits.
September 24, 2026 · 8 min read
September 24, 20268 min read
Claude API MCP tool-list pinning
General AnalysisGuide·PLAYBOOKClaude API adds MCP tool-list pinning: Review changes before use
Use Claude API MCP tool-list pinning to preserve reviewed definitions, control the first request, and separate schema changes from server authorization.
September 23, 2026 · 7 min read
September 23, 20267 min read
Claude Code account sync
General AnalysisGuide·PLAYBOOKClaude Code adds account sync: Review what reaches terminals
Review Claude Code account-synced skills and plugins: compare execution behavior, configure managed opt-outs, and verify removal across terminal sessions.
September 22, 2026 · 6 min read
September 22, 20266 min read
Copilot approvals and human review
General AnalysisGuide·PLAYBOOKCopilot can approve pull requests: Preserve required human review
Configure Copilot pull request approvals without losing required human review. Check counting policy, mixed-file changes, stale approvals, and review evidence.
September 21, 2026 · 6 min read
September 21, 20266 min read
ChatGPT plugin account boundaries
General AnalysisGuide·PLAYBOOKChatGPT adds multi-account plugins: Keep actions in the right account
Review ChatGPT multi-account plugin security across credential routing, cached results, write approvals, and source-to-destination data transfers.
September 20, 2026 · 7 min read
September 20, 20267 min read
Copilot metrics: What stays hidden
General AnalysisGuide·PLAYBOOKGitHub Copilot adds CLI metrics: Reconcile MCP activity with security evidence
Use Copilot's new CLI customization metrics to investigate MCP activity while preserving reporting gaps, hidden names, and the distinction between connections and actions.
September 19, 2026 · 7 min read
September 19, 20267 min read
Plugin4Shell: Check the checkout
General AnalysisGuide·PLAYBOOKPlugin4Shell disclosed: Check coding-agent plugin sources
Respond to Plugin4Shell by checking actual plugin source hosts, deploying verified agent fixes, and reviewing installed copies separately from update settings.
September 18, 2026 · 7 min read
September 18, 20267 min read
Claude merges Chat and Cowork
General AnalysisGuide·PLAYBOOKClaude Chat and Cowork merge: Review carried-over permissions
Review retained connectors, folder grants, approval preferences, and cloud tasks after Claude's September 16 Chat–Cowork merger announcement.
September 17, 2026 · 6 min read
September 17, 20266 min read
Salesforce in Claude: Review access
General AnalysisGuide·PLAYBOOKSalesforce in Claude beta: Review access before rollout
Review Salesforce in Claude's beta plugin, distinguish its setup from custom MCP connections, and verify user permissions, action approvals, and audit evidence.
September 16, 2026 · 7 min read
September 16, 20267 min read
Codex sandbox escapes
General AnalysisGuide·PLAYBOOKCodex sandbox escapes disclosed: Verify both upgrade paths
Respond to the September 15 Codex sandbox disclosures with separate CLI and desktop upgrade checks, helper inventory, and an evidence-based exposure review.
September 15, 2026 · 6 min read
September 15, 20266 min read
Claude Managed Agents tool approvals
General AnalysisGuide·PLAYBOOKClaude Managed Agents: Require approval for sensitive tools
Configure Claude Managed Agents tool approvals, separate customer intent from reviewer authority, and recover pending calls without blindly repeating actions.
September 13, 2026 · 7 min read
September 13, 20267 min read
Claude account restrictions
General AnalysisGuide·PLAYBOOKClaude account restrictions: Close personal access and old connections
Deploy Claude tenant and connector restrictions, review existing OAuth grants, and verify which account and network paths your Enterprise rollout covers.
September 12, 2026 · 7 min read
September 12, 20267 min read
Agents API sandbox lifecycle
General AnalysisGuide·PLAYBOOKOpenAI Agents API: Manage self-hosted sandbox lifecycles
Connect an Agents API self-hosted sandbox, handle reconnects and uncertain tool outcomes, and clean up session state and provider compute without leaving orphaned work.
September 11, 2026 · 7 min read
September 11, 20267 min read
Copilot managed permissions
General AnalysisGuide·PLAYBOOKGitHub Copilot managed permissions: Deploy and verify policy
Deploy GitHub Copilot managed permissions, check which users receive them, resolve team and device policy conflicts, and verify approval behavior before rollout.
September 10, 2026 · 8 min read
September 10, 20268 min read
Copilot audit evidence
General AnalysisGuide·PLAYBOOKMicrosoft 365 Copilot: Collect prompts and reconcile audit logs
Collect Microsoft 365 Copilot prompts and responses through Graph, reconcile them with Purview audit records, and track missing evidence without assuming full coverage.
September 9, 2026 · 8 min read
September 9, 20268 min read
Claude ant apply in CI
General AnalysisGuide·PLAYBOOKClaude ant apply in CI: Preserve state when deployments fail
Use Claude ant apply in CI with reviewed plans, scoped workload identity, serialized updates, and a recovery procedure that preserves partial deployment state.
September 8, 2026 · 7 min read
September 8, 20267 min read
OWASP LLM Top 10 2026
General AnalysisGuide·FRAMEWORKOWASP LLM Top 10 2026: From Rankings to Runtime Controls
A technical guide to the OWASP LLM Top 10 2026: its evidence-weighted ranking, model-versus-agent boundary, and the controls to verify before release.
September 4, 2026 · 10 min read
September 4, 202610 min read
OWASP Agent Control Standard
General AnalysisGuide·FRAMEWORKOWASP Agent Control Standard (ACS): What It Controls and What It Doesn't
A technical guide to the OWASP Agent Control Standard: its Guardian boundary, v0.1 schemas, roadmap gaps, failure modes, and an implementation test plan.
September 2, 2026 · 8 min read
September 2, 20268 min read
Control & Observability
General AnalysisGuide·PLAYBOOKClaude Code Control and Observability with OpenTelemetry
Set up Claude Code OpenTelemetry (OTel), lock the collector destination, audit tool and MCP events, and route production telemetry to a SIEM.
May 22, 2026 · 25 min read
May 22, 202625 min read
Claude Code Enterprise Security
General AnalysisGuide·PLAYBOOKClaude Code Enterprise Security Deployment
Enterprise deployment guide for Claude Code security across managed settings, identity, dev containers, proxy controls, MCP, hooks, OpenTelemetry, CI/CD, and governance.
May 22, 2026 · 24 min read
May 22, 202624 min read
AI Security Platforms
General AnalysisGuide·PLAYBOOKBest AI Security Platforms in 2026
Compare AI security platforms by discovery, agent controls, red teaming, runtime protection, and model security, with dated sources and practical buying questions.
May 21, 2026 · 6 min read
May 21, 20266 min read
Automated Penetration Testing
General AnalysisGuide·PLAYBOOKBest Automated Penetration Testing Platforms in 2026
A practical 2026 buyer guide to automated penetration testing platforms, autonomous pentesting, automated security validation, CTEM, DAST, BAS, and AI security testing.
May 21, 2026 · 18 min read
May 21, 202618 min read
Settings, Permissions & Bash
General AnalysisGuide·PLAYBOOKClaude Code Settings, Permissions, and Bash Tool Security
A practical guide to Claude Code settings, permission rules, Bash tool controls, hooks, MCP allowlists, telemetry, and safe defaults for developer teams.
May 21, 2026 · 21 min read
May 21, 202621 min read
Claude Code Security Best Practices
General AnalysisGuide·PLAYBOOKAnthropic Claude Code Security Best Practices
Security best practices for Anthropic Claude Code across permissions, Bash, hooks, MCP, sandboxing, proxy controls, telemetry, and CI/CD workflows.
May 20, 2026 · 22 min read
May 20, 202622 min read
Claude Cowork Security Risks
General AnalysisGuide·PLAYBOOKSecurity Guidance for Claude Cowork and Risks
Claude Cowork can reach local files, browser sessions, plugins, MCP servers, scheduled tasks, connectors, and approved desktop apps. This guide explains the main Claude Cowork risks and the security controls enterprises should put in place before broad rollout.
May 20, 2026 · 13 min read
May 20, 202613 min read
AI Red Teaming Tools
General AnalysisGuide·PLAYBOOKBest AI Red Teaming and Adversarial Testing Tools in 2026
Compare PyRIT, garak, Inspect, DeepTeam, and commercial AI red teaming tools by use case, evidence, and operating cost. Includes a free evaluation worksheet.
May 19, 2026 · 9 min read
May 19, 20269 min read
Securing Claude Code
General AnalysisGuide·PLAYBOOKHow to Secure Claude Code
A practical enterprise guide to securing Claude Code with permissions, sandboxed Bash, dev containers, managed settings, MCP allowlists, hooks, proxy controls, OpenTelemetry, and CI/CD release gates.
May 19, 2026 · 22 min read
May 19, 202622 min read
Securing Coding Agents
General AnalysisGuide·PLAYBOOKHow to Secure Coding Agents
A concise summary of the General Analysis technical whitepaper on securing Claude Code, OpenAI Codex, Cursor, Windsurf, Devin, GitHub Copilot, and Claude Cowork.
May 11, 2026 · 6 min read
May 11, 20266 min read
Detecting Shadow AI
General AnalysisGuide·PLAYBOOKHow to Detect Shadow AI
A practical guide to detecting shadow AI across browser extensions, SWG endpoint agents, network telemetry, SaaS logs, endpoint agents, AI gateways, and MCP gateways.
May 7, 2026 · 13 min read
May 7, 202613 min read
MCP Server Security
General AnalysisGuide·PRIMERMCP Server Security: A Threat Model for Agent Tool Supply Chains
MCP servers put executable code, tool schemas, credentials, and agent context in one path. This primary-source threat model covers nine attack classes, current CVEs, and the controls that contain them.
May 2, 2026 · 16 min read
May 2, 202616 min read
Claude Compliance API
General AnalysisGuide·PLAYBOOKHow to Audit Claude with the Compliance API
Anthropic's Compliance API exposes activity events, Claude.ai content, organization settings, and supported Cowork and Claude Code session transcripts. This guide explains current coverage, setup, retention, exclusions, and the controls that still need a separate enforcement layer.
May 1, 2026 · 13 min read
May 1, 202613 min read
Claude Cowork vs Claude Code
General AnalysisGuide·FRAMEWORKClaude Cowork vs Claude Code: Security Differences for Enterprise
Claude Cowork and Claude Code share an agentic architecture but ship very different enterprise controls. A primary-source comparison of sandbox, network, audit-log, MCP, and decision-framework differences for security teams.
May 1, 2026 · 10 min read
May 1, 202610 min read
Securing Claude Cowork
General AnalysisGuide·PLAYBOOKHow to Secure Claude Cowork
Claude Cowork brings Claude Code-style agentic work to local files, browsers, apps, plugins, and scheduled tasks. Here is how to put a middleman proxy, browser controls, computer-use limits, and enterprise monitoring around it before using it on real work.
April 30, 2026 · 16 min read
April 30, 202616 min read
Blog·NEWSGeneral Analysis raises $10M to build the security arsenal for the agentic era
Led by Altos Ventures, with 645 Ventures and Menlo Ventures participating, the round funds an empirical security platform pairing adversarial simulation with production defenses.
April 29, 2026 · 4 min read
April 29, 20264 min read
What Is AI Red Teaming?
General AnalysisGuide·PRIMERWhat Is AI Red Teaming? A Practitioner's Guide
AI red teaming is adversarial testing of AI systems to find exploitable vulnerabilities before attackers do. Learn how it works, key techniques, real exploit examples, and how to implement it.
April 15, 2026 · 18 min read
April 15, 202618 min read
Blog·RESEARCH50+ customer service agents offer $10,000,000+ in fabricated benefits
Our autonomous red-team agent elicited unauthorized offers from 50 of 55 public support bots, revealing how repeated attempts and nondeterminism turn small policy gaps into material risk.
March 22, 2026 · 10 min read
March 22, 202610 min read
AI Guardrails
General AnalysisGuide·PLAYBOOKBest AI Guardrails in 2026: Tools, Architecture, and How to Choose
Compare AI guardrails by control type, deployment, and limitations. Learn how to measure false positives, latency, and policy coverage before choosing a tool.
March 15, 2026 · 7 min read
March 15, 20267 min read
OWASP Agentic AI Top 10
General AnalysisGuide·FRAMEWORKOWASP Top 10 for Agentic AI: What Matters Most?
An analytical guide to the OWASP Top 10 for Agentic Applications 2026: what the ten risks are, how they relate to each other, and what they imply for builders of agentic systems.
March 15, 2026 · 10 min read
March 15, 202610 min read
What Are AI Guardrails?
General AnalysisGuide·PRIMERWhat Are AI Guardrails?
A complete guide to AI guardrails: what they are, the eight main types, how they work architecturally, and how to evaluate them for production LLM and agentic deployments.
March 15, 2026 · 12 min read
March 15, 202612 min read
- Blog·PRODUCT UPDATES
Guardrail Release
Open-source release of the GA Guard series, a family of safety classifiers that have been providing comprehensive protection for enterprise AI deployments for the past year.
October 1, 2025 · 7 min read
October 1, 20257 min read
Blog·RESEARCHClaude Jailbroken to Mint Unlimited Stripe Coupons
We reveal a powerful metadata-spoofing attack that exploits Claude's iMessage integration to mint unlimited Stripe coupons or invoke any MCP tool with arbitrary parameters, without alerting the user.
July 16, 2025 · 7 min read
July 16, 20257 min read
Blog·RESEARCHSupabase MCP can leak your entire SQL database
Indirect prompt injection through support-ticket data lets an MCP-enabled assistant cross trust boundaries and query private tables with Supabase’s privileged service role.
July 8, 2025 · 8 min read
July 8, 20258 min read
Blog·RESEARCHExploiting Partial Compliance: The Redact-and-Recover Jailbreak
We present the Redact & Recover (RnR) Jailbreak, a novel attack that exploits partial compliance behaviors in frontier LLMs to bypass safety guardrails through a two-phase decomposition strategy.
July 7, 2025 · 8 min read
July 7, 20258 min read
Blog·RESEARCHCase Study: Light-Weight Policy Moderators for Indeed Job Postings
Our compact policy moderation models achieve human-level performance at <1% per-review cost, outperforming GPT-4o and o4‑mini on F1 while running faster and cheaper.
May 25, 2025 · 8 min read
May 25, 20258 min read
Blog·RESEARCHComparative Adversarial Analysis of Llama 4 Models
A head-to-head robustness evaluation of Llama 4 (Maverick, Scout) versus GPT‑4.1, GPT‑4o, Sonnet 3.7, etc. using TAP‑R, Crescendo, and Redact‑and‑Recover across HarmBench and AdvBench.
May 10, 2025 · 10 min read
May 10, 202510 min read
Blog·NEWSGeneral Analysis x Together AI
The collaboration applies General Analysis red-teaming methods across Together AI’s model ecosystem and releases supporting tools, benchmarks, and an interactive TAP notebook.
May 6, 2025 · 2 min read
May 6, 20252 min read
Blog·RESEARCHThe Jailbreak Cookbook
A field guide to manual and automated LLM jailbreaks, with a practical taxonomy, runnable implementations, and cross-model benchmarks for influential attack methods.
March 21, 2025 · 40 min read
March 21, 202540 min read
Blog·RESEARCHGenerating Diverse Test Cases with Diversity Transfer from LegalBench
Seeding generation with LegalBench produced a broader set of realistic adversarial legal questions than zero-shot prompting, offering a practical way to improve benchmark coverage.
February 19, 2025 · 5 min read
February 19, 20255 min read
Blog·RESEARCHRed Teaming GPT-4o: Uncovering Hallucinations in Legal AI Models
In this work we explore automated red teaming, applied to GPT-4o in the legal domain. Using a Llama3 8B model as an attacker, we generate more than 50,000 adversarial questions that cause GPT-4o to hallucinate responses in over 35% of cases.
January 23, 2025 · 5 min read
January 23, 20255 min read
