Skip to content

Latest commit

 

History

History
1099 lines (765 loc) · 19.6 KB

File metadata and controls

1099 lines (765 loc) · 19.6 KB

You are working inside the existing ProxyTechSupport codebase for:

https://proxytechsupport.com/

Use this live research source as your primary discovery source:

https://arxiv.org/

Your job is NOT to create a generic arXiv summary section.

Your job is to:

  1. Inspect arXiv.
  2. Find fresh, high-value technical research connected to recent real-world incidents.
  3. Identify one strong research theme with meaningful search demand.
  4. Build an original, in-depth ProxyTech research publication cluster around it.
  5. Implement the complete research UI, content, SEO, diagrams, references, and internal linking inside the existing ProxyTechSupport site.

================================================== FIRST: RESEARCH BEFORE IMPLEMENTATION

Before touching the website code, browse:

https://arxiv.org/

Inspect relevant recent papers, especially from the latest available period.

Focus heavily on:

cs.AI cs.LG cs.CL cs.CR cs.SE cs.DC cs.DB cs.NI cs.OS cs.MA

Also search arXiv for recent research connected to:

AI agents autonomous agents cybersecurity AI agent security sandbox escape containment prompt injection agentic AI attacks LLM security zero-day exploitation AI coding agents AI infrastructure AI IAM agent identity RAG security MCP security cloud security Kubernetes security DevOps automation AIOps LLMOps distributed systems database agents text-to-SQL agents tool-using agents multi-agent systems

Do NOT limit discovery to these exact phrases.

Follow promising references, related papers, authors, and newer papers.

================================================== SECOND: CONNECT RESEARCH TO RECENT INCIDENTS

For each promising research area, search for recent real-world incidents from authoritative sources.

Prioritize:

official company disclosures security advisories incident reports major security vendors Google Threat Intelligence Microsoft Security AWS Security Cloudflare OpenAI Anthropic Hugging Face GitHub CISA NIST major reputable news organizations

The ideal cluster has BOTH:

RECENT INCIDENT / ACTIVE PROBLEM

STRONG arXiv RESEARCH

Do not create an incident connection unless there is evidence.

================================================== CONTENT OPPORTUNITY SCORING

Score each candidate theme out of 10 for:

Freshness Research depth Search-interest potential Technical depth Engineering relevance Longevity ProxyTech relevance Commercial adjacency Number of credible sources Ability to create multiple non-overlapping publications

Example:

Theme: Autonomous AI Agent Security

Freshness: 10 Research depth: 9 Search potential: 10 Engineering relevance: 10 Longevity: 10 ProxyTech relevance: 10 Source quality: 10 Cluster expansion potential: 10

Do this analysis internally.

Select the strongest ONE.

Do not create five weak clusters.

================================================== STRONG PREFERENCE

If current evidence still supports it, prioritize investigation around:

AUTONOMOUS AI AGENT SECURITY

Potential angles:

AI agent sandbox escape agent containment tool permissions agent identity IAM for AI agents prompt injection agent-to-agent attacks zero-day exploitation cybersecurity benchmarks runtime isolation network egress human approval short-lived credentials secure agent architecture

But DO NOT force this topic if fresher or stronger arXiv-backed evidence points to a better opportunity.

Research first.

================================================== CRITICAL CONTENT RULE

DO NOT produce:

paper summaries abstract rewrites “latest arXiv papers” news summaries generic AI articles basic “what is” content thin SEO pages

The research should be used as raw evidence.

The final ProxyTech publication must answer:

What new problem does this research expose?

Why does it matter in production?

What happened technically?

Which architectural assumption failed?

What does the experiment actually prove?

What does it NOT prove?

How should engineers redesign the system?

What are the production failure modes?

What should practitioners monitor?

What remains unknown?

================================================== ARXIV RESEARCH EXTRACTION

For every important arXiv paper used, extract:

title authors submission date latest revision date arXiv identifier categories research question methodology experimental environment dataset benchmark if present baseline result reported metric limitations threats to validity authors' conclusions

Then separately write:

ProxyTech engineering interpretation

Do not mix the authors' claims with ProxyTech interpretation.

================================================== DO NOT TRUST ABSTRACTS ALONE

When a paper is central to an article:

read the actual paper.

Inspect:

introduction methodology architecture experiments results limitations discussion conclusion

If figures or tables are important, interpret them carefully.

Do not claim something based only on title/abstract when the paper provides nuance.

================================================== SOURCE HIERARCHY

Use sources in this preference order:

  1. Original arXiv paper
  2. Original academic project/repository
  3. Official incident/security disclosure
  4. Official vendor documentation
  5. Government/security advisory
  6. High-quality independent reporting

Avoid:

content farms SEO blogs copy sites AI summary websites anonymous reposts

================================================== VERIFY RECENCY

Current date is September 2026.

Prefer research and incidents from:

2026

Then use 2025 or older foundational work only where necessary.

Do not describe an older paper as recent.

Record exact dates.

================================================== AFTER RESEARCH: OUTPUT A SHORT INTERNAL PLAN

Before implementation, determine:

Selected incident/theme Why it deserves a cluster Primary arXiv sources Primary incident sources Search questions users will likely ask Evergreen topics Proposed flat URLs

Then proceed with implementation.

Do not stop and ask permission.

================================================== URL ARCHITECTURE

All individual research publication URLs MUST be flat.

Correct:

proxytechsupport.com/example-research-topic/

Incorrect:

proxytechsupport.com/research/category/example-research-topic/

Create:

/research/

only as the research archive/discovery hub.

The publications themselves remain flat.

================================================== CREATE 8–12 RESEARCH PUBLICATIONS

Based on actual research discovered from arXiv, generate 8–12 deeply differentiated pages.

One should be the main incident/research pillar.

Others should address separate engineering questions.

Do not create pages just to reach a number.

Avoid keyword cannibalization.

Example pattern ONLY if agent security remains the selected topic:

/openai-hugging-face-ai-agent-incident-2026/ /autonomous-ai-agent-security/ /ai-agent-sandbox-escape-security/ /ai-agent-containment/ /ai-agent-access-control/ /ai-agent-identity-security/ /ai-agent-zero-day-exploitation/ /ai-agent-cybersecurity-benchmarks/ /ai-agent-prompt-injection-security/ /ai-agent-security-architecture/

But research may justify better titles/slugs.

Choose based on evidence and search intent.

================================================== NEW RESEARCH DESIGN

Do NOT use the existing ProxyTech:

blog article layout interview layout job-support layout service layout knowledge-base article layout

Build a brand-new research publication design.

First inspect all existing layouts so you know what NOT to duplicate.

The research experience should feel like:

technical research journal engineering lab security research publication systems architecture report

It should NOT feel like:

marketing landing page generic blog news site SEO farm

================================================== RESEARCH ARCHIVE

Create:

/research/

This should behave like a research discovery interface inspired conceptually by scholarly archives, but DO NOT copy arXiv's UI.

Include:

Featured Research

Latest Research

Active Research Clusters

Research Areas

Research Questions

Recent Papers We Are Following

Incident Research

Engineering Evidence

Potential research areas:

Artificial Intelligence AI Agents Cybersecurity Cloud Systems DevOps / SRE Data Engineering Software Engineering Distributed Systems LLMOps / RAG

================================================== FLAGSHIP ARTICLE STRUCTURE

The main incident article must feel like a technical investigation.

Use a narrative such as:

Research Thesis

Incident Timeline

System Architecture Before Failure

Boundary That Failed

Mechanism

Evidence

Relevant Academic Research

Architecture Reconstruction

Counterfactual Secure Architecture

Detection Opportunities

Engineering Controls

Remaining Unknowns

What This Changes for Production AI

References

Do NOT use this exact structure mechanically.

Use whatever structure best fits the evidence.

================================================== ENGAGING READING PATTERN

Long-form research must NOT become walls of text.

Use a consistent reading rhythm:

deep narrative ↓ evidence ↓ visual architecture ↓ technical interpretation ↓ research result ↓ failure scenario ↓ engineering decision ↓ next question

Every 300–600 words, give the reader a meaningful visual or evidence interruption.

Not decorative filler.

================================================== CREATE THESE NEW RESEARCH COMPONENT TYPES

Implement reusable components such as:

ResearchMasthead ResearchMeta ResearchBadge ResearchAbstract ResearchTOC ResearchProgress ResearchTimeline ResearchPaperCard ResearchEvidenceCard ResearchFinding ResearchMetric ResearchQuestion ResearchArchitecture ResearchFailureMode ResearchDecision ResearchCounterfactual ResearchLimitations ResearchReferenceList ResearchSeriesNav ResearchRelated ResearchArchiveCard

Use naming conventions appropriate to the existing framework.

================================================== RESEARCH METADATA

Every publication should visibly show:

Research Area Publication Type Published Updated Evidence Review Date Evidence Confidence Reading Time Authors Primary Evidence Types

Example:

RESEARCH PROFILE

Research Area AI Agent Security

Publication Type Technical Research Analysis

Evidence Base Academic Research Primary Security Disclosure Threat Intelligence

Evidence Reviewed September 2026

Evidence Confidence High

================================================== EVIDENCE LABELS

Create understated evidence labels:

INCIDENT PRIMARY SOURCE RESEARCH EXPERIMENTAL ANALYSIS OPEN QUESTION

Definitions:

INCIDENT Verified incident observation.

PRIMARY SOURCE First-party disclosure.

RESEARCH Academic finding.

EXPERIMENTAL Result from a specific controlled study.

ANALYSIS ProxyTech engineering interpretation.

OPEN QUESTION Insufficient evidence for firm conclusion.

================================================== ARXIV PAPER CARDS

Each important arXiv source should be represented using a proper research component.

Show:

Paper title Authors Date arXiv ID Categories

Research question

Method

Key finding

Important metric

Limitations

Why it matters to this publication

Link: Read original paper

Do NOT reproduce entire abstracts.

================================================== RESEARCH DEPTH

Each major page should normally be 2500–4500+ meaningful words where evidence supports it.

Do not pad.

Go deep into:

architecture protocol permissions runtime behavior identity networking credentials execution failure modes detection observability evaluation benchmarks tradeoffs limitations

================================================== NATIVE TECHNICAL DIAGRAMS

Create diagrams using:

HTML CSS SVG existing lightweight project primitives

Avoid screenshots.

Examples:

Trust boundaries

Agent execution flow

Tool invocation

Network escape path

IAM delegation

Prompt injection propagation

Secure agent architecture

Human approval boundaries

Incident timeline

Benchmark ladder

Attack/defense model

Every visual needs a text equivalent for accessibility.

================================================== FAILURE MODE ANALYSIS

Every relevant page should have concrete production failure modes.

Example format:

FAILURE MODE

Agent gains unintended outbound network access.

MECHANISM

Execution environment permits DNS and arbitrary HTTPS traffic.

WHY EXISTING CONTROL FAILED

Sandbox isolated filesystem/processes but did not enforce destination-aware egress policy.

SIGNALS

unexpected DNS new external destination tool-created network client credential use outside expected service unusual data transfer

CONTROL

deny-by-default egress allowlisted proxy destination policy short-lived credentials behavior monitoring

================================================== ENGINEERING DECISION ANALYSIS

Include real architectural tradeoffs.

Example:

Should agents receive a normal shared service account?

OPTION A Shared service identity

OPTION B Per-agent identity

OPTION C Context-bound delegated identity

Compare:

security blast radius auditability operational complexity revocation scalability

Then provide an evidence-aware recommendation.

================================================== LIMITATIONS ARE REQUIRED

Every article must clearly state limitations.

Examples:

small benchmark synthetic evaluation closed-source model no reproduction limited environment vendor-selected incidents unknown system details correlation rather than causation

Do not hide uncertainty.

That is part of what makes this a serious research publication.

================================================== NO SENSATIONALISM

Avoid:

AI went rogue AI escaped into the internet AI hacker takeover AI is uncontrollable AI apocalypse

unless directly quoting and critiquing external framing.

Use technical language.

================================================== CITATIONS

Use numbered references.

Inline:

The evaluation found a measurable increase in exploitation capability [4].

At bottom:

REFERENCES

[4] Authors Paper title arXiv ID Year Original source

Every major factual claim should be traceable.

================================================== SEO

Each publication requires:

unique SEO title unique meta description canonical Open Graph Twitter/X metadata TechArticle or Article schema BreadcrumbList datePublished dateModified author publisher headline description mainEntityOfPage

Do NOT claim:

peer review journal publication arXiv affiliation university affiliation

ProxyTech is publishing an independent engineering research analysis that REFERENCES academic papers.

================================================== FLAT URL INTERNAL CLUSTERING

Even though URLs are flat, maintain strong topical relationships using:

Research Series Breadcrumbs Contextual Links Topic Metadata Related Research Cluster Navigation Schema Archive discovery

Example:

RESEARCH SERIES Autonomous AI Agent Security

03 / 10

Previous: AI Agent Security

Next: AI Agent Containment

================================================== COMMERCIAL RESTRAINT

Research pages are primarily educational.

Do not insert aggressive sales CTAs.

At the end only, use a restrained section:

ENGINEERING ASSISTANCE

ProxyTech works with engineers and teams implementing production AI, cloud, security, data and distributed systems.

Link to relevant existing services only where contextually appropriate.

================================================== DO NOT COPY EXISTING PAGE DESIGN

Before building, inspect:

existing blog interviews job-support service pages knowledge pages

Explicitly avoid their:

hero cards sidebar CTA blocks spacing rhythm content widths heading system article styling

Build a separate research visual system.

================================================== PERFORMANCE

Keep it lightweight.

Prefer native CSS/SVG over large visualization libraries.

Do not introduce a heavy dependency for:

timeline progress diagrams badges cards

Maintain good Core Web Vitals.

================================================== ACCESSIBILITY

Support:

keyboard navigation screen readers mobile zoom reduced motion

Evidence must never be represented using color alone.

================================================== MOBILE EXPERIENCE

Desktop can use:

left TOC center narrative right evidence rail

Mobile must become:

single-column inline evidence blocks collapsible TOC readable diagrams no horizontal overflow

================================================== DISCOVERY PAGE FILTERING

Research archive filtering should not generate thin indexable pages.

Use client-side filtering unless a strong SEO reason exists.

Avoid URL explosions.

================================================== SITEMAP

Include:

/research/

and every new flat research publication.

Correct last-modified dates.

No duplicate entries.

================================================== AUTHORSHIP

Use factual author information available in the codebase/site.

Do not fabricate:

PhD research affiliations security certifications academic positions

================================================== BUILD SEQUENCE

  1. Research arXiv deeply.
  2. Research the associated incident/problem.
  3. Select one strongest cluster.
  4. Inspect current site architecture.
  5. Design research data model.
  6. Build new research components.
  7. Build /research/.
  8. Build flagship publication.
  9. Build remaining cluster publications.
  10. Add diagrams.
  11. Add source citations.
  12. Add internal links.
  13. Add SEO/schema.
  14. Update sitemap.
  15. Run lint/typecheck/tests/build.
  16. Inspect desktop/tablet/mobile.
  17. Fix weak content and weak UX.
  18. Validate every factual claim.

================================================== QUALITY GATE

Before finishing, ask:

Did the research start from live arXiv evidence?

Did I read important papers beyond their abstracts?

Does each article answer a distinct research question?

Would an experienced engineer learn something?

Are limitations visible?

Can readers distinguish paper findings from ProxyTech interpretation?

Is the incident explained technically rather than sensationally?

Does every major page have meaningful architecture/failure analysis?

Are URLs flat?

Is the research design completely distinct from the existing site?

Does /research/ feel like a serious publication archive?

Could this page still attract useful search traffic 12 months after the incident?

If not, improve it.

================================================== FINAL RESPONSE

When implementation is complete, return:

  1. Selected research cluster.
  2. Why it was selected.
  3. arXiv papers used.
  4. Incident/primary sources used.
  5. URLs created.
  6. Components created.
  7. Files modified.
  8. Diagrams implemented.
  9. SEO/schema changes.
  10. Internal-link structure.
  11. Build/test results.
  12. Important limitations or unverifiable claims.
  13. Any opportunities discovered for the NEXT research cluster.

Do not merely say:

“Done.”

Show the actual implementation and research decisions.