Context Jamming · A reporting hub
AI × Cybersecurity
Discovery is nearly free now; the scarce thing in security is the fix. The harness around a model matters more than the model, and the perimeter still faces outward at attackers who already hold the keys.
37 pieces · 4 threads · 10 months of reporting since January 2026
Start here

Article· Substack (opens in a new tab)
The Perimeter Faces the Wrong Way
AI security has two front doors and the perimeter faces neither. Discovery is nearly free now; the scarce asset is the fix.

Dispatch
The Harness Is the Weapon
The first self-sustaining AI worm runs on a 2025 open-weight model and one A100. The scaffolding is the weapon, not the model.

Article· Substack (opens in a new tab)
The Twenty-Seven-Year Bug
One quarter forced AI to decide what it was for: a 27-year-old bug, Mythos, Glasswing, and the day defense became the moat.
Interactive
The Classifier Watches the Exchange
Anthropic’s jailbreak defense moved into the generation stream and got about 40× cheaper by watching the whole exchange.
The reporting
Showing 37 of 37
How AI is changing security
Discovery got cheap, agents got credentials, and the harness became the weapon.

Article· Substack (opens in a new tab)· Context
Poison, Gates and the Price of a Pirated Library
Creators won leverage through gates, contracts and courts, not by proving that data poisoning corrupts frontier models.

Article· Substack (opens in a new tab)
The Perimeter Faces the Wrong Way
AI security has two front doors and the perimeter faces neither. Discovery is nearly free now; the scarce asset is the fix.

Article· Substack (opens in a new tab)· Context
The Screener Is Both the Attack Surface and the Defense
AI hiring screeners can’t tell an instruction from an answer, or a rendered face from a real one. Prompt injection lands in HR.

Dispatch
The Harness Is the Weapon
The first self-sustaining AI worm runs on a 2025 open-weight model and one A100. The scaffolding is the weapon, not the model.

Interactive
The Cybersecurity CMO’s AI Spend Approval Map
A GTM operating brief for SaaS security CMOs: GEO visibility, AI-claim governance, trust architecture, and where AI budget actually gets approved.

Article· Substack (opens in a new tab)· Context
Your AI Agent Is Leaking Intelligence
Field notes from AgentCon Boston on how agents leak what they know.
Interactive
Tokenization Consensus Layer
An execution firewall that flags adversarial input where frontier tokenizers disagree. Live instrument, math included.
Dispatch
The Scaffolding Is the Product
Stanford’s ARTEMIS beat 9 of 10 human pentesters at $18.21 an hour. Same model, three scaffolds, a 2.7× gap.

Article· Substack (opens in a new tab)
The Sentinel and the Swarm
MCP became the USB-C of intelligence and rewired the attack surface. The 2026 enterprise wasn’t hacked by a kid in a hoodie.
Anthropic's cyber footprint
Glasswing, Mythos, classifiers, red-teaming research, and the lab’s own incidents.

Article· Substack (opens in a new tab)· Context
The Box on Line 10
A lobbying line, a Rubio memo and the Mythos export fight: four days of posts with no filing behind them.

Video
Guard Inside the Castle
Constitutional Classifiers++ as a pop-up book: one guard who reads the question and the answer together.

Video· Context
Demon at the Gate
Fifteen seconds from the semantic-entropy classifier explainer: the demon sorts the stream.

Video
Subtle Misalignment Arcade
A vintage-arcade walkthrough of how Anthropic automated red-teaming for subtle emergent misalignment.

Article· Context
The Audited Fact
The S-1 read against the book: the state shows up as a risk factor, and Glasswing shows up in the filing.

Article· Substack (opens in a new tab)· Context
The Six People Inside the Perimeter
Embedded evaluators are how Anthropic tries to bridge secrecy and trust without hiding the economics of inference.

Article· Substack (opens in a new tab)
Three Days of Daylight
Anthropic’s fabricated police tip took 81 days to reach Philadelphia; OpenAI’s Australia breach took 84. Fast once found, slow to find.

Article· Substack (opens in a new tab)· Context
Amodei’s Guardrails Became the Soft-Nationalization Playbook
The case for central oversight became the architecture for state control of frontier AI. The export-control thread.

FounderFile· Context
Tom Brown · FounderFiles N°046
The compute diplomat who unlocked Mythos and Fable, and brokered the access that came with them.

Article· Context
The Grounding
June 12, 2026: Washington put Fable 5 and Mythos 5 under export controls. The Anthropic Book, chapter 14.

Article· Substack (opens in a new tab)
Anthropic’s Mythos Gambit Just Hit the Jagged Frontier
A 244-page system card, record SWE-bench, and a 3.6B model that replicated the flagship zero-day. Tiered access meets the jagged frontier.

Article· Substack (opens in a new tab)· Context
The Glass Box and the Orchestration of the Wolf
May 2026: the month the industry slipped quietly into a new threat reality.

Article· Substack (opens in a new tab)
The Twenty-Seven-Year Bug
One quarter forced AI to decide what it was for: a 27-year-old bug, Mythos, Glasswing, and the day defense became the moat.

Article· Substack (opens in a new tab)· Context
Deep Research as a Weapon
The layer the Anthropic–Pentagon fight is missing.

Article· Substack (opens in a new tab)· Context
The Holographic Leak
59.8 MB of forgotten code, read for the physicist’s fingerprint inside the world’s most powerful agent.

Article· Substack (opens in a new tab)
The Leaked Claude Code Blueprint
Anthropic shipped all 512k lines of Claude Code through a forgotten npm source map. What the blueprint gives away.

Article· Substack (opens in a new tab)
The “++” Signal: When Safety Becomes a Commodity
Classifiers++ marks the Great Internalization: safety becomes a feature the labs ship, and security vendors feel the squeeze.

Infographic· Substack (opens in a new tab)
Claude Code Security: A Five-Point Strategic Audit
Five points on what Claude Code Security means for the vulnerability market, drawn as one Claude-made infographic.

Article· Substack (opens in a new tab)
The Sentinel Protocol
The earliest read in the run: Anthropic was walking into autonomous cyber defense before anyone priced it in.
Interactive
The Classifier Watches the Exchange
Anthropic’s jailbreak defense moved into the generation stream and got about 40× cheaper by watching the whole exchange.
OpenAI and the other labs
What it means when the model companies start selling security.
From the field
A decade at Mimecast, a quarter with Proofpoint, and the people who built email security.

Case study
The Mimecast Studio
The video studio I built at Mimecast in 2016 and the Lexington news set, from SKO stages to the Black Hat floor.

FounderFile
Evan Reiser · FounderFiles N°040
Abnormal’s founder turned ad-tech behavioral prediction into an API-native email-security primitive.

Case study
ACRA Insight: The Proofpoint Engagement
One quarter of AI GTM consulting for Proofpoint’s CMO org, written up in public.

FounderFile· Context
David Raissipour · FounderFiles N°016
Builds aggregation layers that turn scattered components into a secure, observable control plane.

FounderFile· Context
Dino DiMarino · FounderFiles N°015
The go-to-market operator who bridges developer-first realities with enterprise risk quantification.

FounderFile
Peter Bauer · FounderFiles N°012
Mimecast’s founder bet two decades on SMTP as unowned infrastructure. Includes the SolarWinds disclosure fracture.

FounderFile· Context
Jeff Schumann · FounderFiles N°011
Aware’s co-founder went from adversarial signal intelligence to governing the enterprise’s human-metadata perimeter.
From the field
Mimecast → Proofpoint → Context Jamming
I spent a decade in email security at Mimecast, from the IPO through the take-private. I built the company’s first video studio in 2016 and later designed the Lexington news set.
In spring 2026 I spent a quarter consulting on AI go-to-market for Proofpoint’s CMO org. The public write-up is on the ACRA Insight page.
Since then I’ve been reporting here on what frontier AI does to the security market: who finds the bugs, who ships the fix, and who gets paid.
Need this read on your market?
I help security companies turn what the labs are shipping into a position their buyers believe: executive narrative, competitive reads and content systems, built with Claude.
