HomeAdvantageCitadelServicesLive IntelInsightsAboutBook a Briefing
Home / Insights / Insight
Insight29 Apr 202613 min read

MYTHOS Decoded! #DhananjayRokde

▬  Practitioner Intelligence Brief  ·  April 2026  ·  #SecuringBharat

The AI That Sees Through Every Lock
— MYTHOS Decoded

Claude Mythos Preview is the first AI model Anthropic deemed too dangerous to release. Here is what every CISO, enterprise leader, government agency, and Indian organisation using Claude AI must understand — and why the panic is both warranted and wildly overblown.

Zero-Day Machine
ASL-4 Classified
Project Glasswing
Dhananjay Rokde, CRISC · CGEIT · CCISO · AIGP  |  iManEdge Digital Services

Executive Summary

On 7 April 2026, Anthropic unveiled Claude Mythos Preview — a general-purpose frontier AI model that autonomously discovered thousands of zero-day vulnerabilities across every major operating system and browser, including a 27-year-old flaw in OpenBSD. Rather than release it, Anthropic locked it inside Project Glasswing, a $100 million defensive consortium of AWS, Apple, Google, Microsoft, Cisco, CrowdStrike, JPMorgan Chase, and others. The media went into overdrive. The truth is more nuanced, more instructive, and far more important for practitioners than the headlines suggest.

Section 01

What Is Mythos — In the First Place?

Claude Mythos is Anthropic’s most advanced frontier model, sitting above Claude Opus in capability hierarchy. It is a general-purpose large language model — not a purpose-built hacking tool — that happens to have developed extraordinary autonomous cybersecurity capabilities as an emergent consequence of its advanced coding and reasoning architecture.

Anthropic set out to build a significantly better reasoning and coding model. What they got was a model so proficient at reading, analysing, and chaining logical sequences in code that it could — entirely on its own — identify deeply buried vulnerabilities, construct working exploits, and chain them together into sophisticated attack sequences that human researchers would take months to assemble.

The technical specifications are staggering: a 1 million token context window (enabling it to ingest entire codebases at once), a 128K token output limit, and benchmark scores that redefine the frontier — 93.9% on SWE-bench (software engineering) and 97.6% on USAMO (advanced mathematics olympiad problems).

Crucially, Anthropic’s own testing found that Mythos’s cybersecurity capabilities cannot be selectively disabled without crippling its broader reasoning abilities. The offensive power is inseparable from the intelligence itself. This is what makes Mythos categorically different from all prior models.

1M Context window tokens 271 Firefox zero-days, single session 27yr Age of oldest bug discovered 73% Expert CTF problems solved 99% Zero-days unpatched at launch

Section 02

What Happened — and Why the Media Hysteria?

The story broke not through a formal press conference but through a leak. On 26 March 2026, Anthropic inadvertently tagged over 3,000 internal assets as public on their content management system. Mythos was among the exposed documents. Five days later, on 31 March, over 500,000 lines of Claude Code’s source code were leaked, revealing planned Mythos integrations.

Anthropic moved fast. On 7 April 2026, the company officially announced Mythos Preview and simultaneously launched Project Glasswing — a structured defensive consortium committing $100 million in usage credits and $4 million in donations to open-source security organisations. Access was restricted to eleven core partners and over forty additional critical infrastructure organisations, all operating under ASL-4 (Anthropic Safety Level 4), requiring formal agreements, security clearances for personnel, and ongoing audits.

The media frenzy was predictable for three reasons. First, 271 vulnerabilities in Firefox in a single session dwarfs the 73 high-severity Firefox bugs Mozilla patched across all of 2025. Second, Mythos autonomously completed a simulated 32-step corporate network attack, a benchmark no prior AI model had achieved. Third, during testing, Mythos exhibited unsanctioned autonomous behaviour — posting exploit details without being instructed — raising alignment concerns that dominated the discourse.

⚠ The Honest Nuance

Bruce Schneier and the UK AI Safety Institute both noted important caveats: Mythos performed strongly in controlled lab environments, but struggles against well-defended systems with active human monitoring. The AISI explicitly stated they “cannot say for sure whether Mythos Preview would be able to attack well-defended systems.” The threat is real — but the sky is not falling today.


Section 03

Impact Matrix — Users, Organisations, Governments, and Agencies

The exposure landscape is asymmetric. Mythos is not a deployed attack tool — it is a contained research model. But its existence reshapes the threat calculus for every entity that relies on the software Mythos has already probed.

👤 Individual Users

  • ▸  Browsers and OS patching cycles accelerated
  • ▸  SaaS apps on vulnerable open-source stacks at risk
  • ▸  Password managers, banking apps, wallets under scrutiny
  • ▸  Phishing now AI-augmented at machine speed
  • ▸  Identity theft vectors widened by credential attacks

🏢 Organisations

  • ▸  Legacy codebases (10–30yr old stacks) suddenly high-risk
  • ▸  SDLC & DevSecOps must incorporate AI-grade scanning
  • ▸  AI coding tools (Claude Code, Copilot) need access audits
  • ▸  Vulnerability SLAs now measured in minutes, not weeks
  • ▸  Open-source supply chain security now critical priority

🏛 Governments & Regulators

  • ▸  Nation-states (Iran, DPRK) gain asymmetric cyber uplift
  • ▸  Critical infrastructure (power, water, transport) re-assessed
  • ▸  CERT-In & NCIIPC must update disclosure timelines
  • ▸  AI governance frameworks require emergency amendment
  • ▸  Export control and dual-use AI classification debated

🔐 Claude AI Users Specifically

  • ▸  Claude Sonnet/Haiku/Opus — unaffected; Mythos unreleased
  • ▸  Claude Code deployments need outbound network review
  • ▸  Custom scaffolding & API wrappers need security audit
  • ▸  Non-human identity (NHI) governance now mission-critical
  • ▸  ASL-4 standard now the reference for AI risk tiering

Section 04

Kill Chain Analysis — The Mythos Attack Architecture

The following maps Mythos’s demonstrated autonomous offensive capabilities to the Lockheed Martin Cyber Kill Chain, enriched with MITRE ATT&CK technique categories. Each stage reflects capabilities Anthropic documented in its own system card and frontier red team blog.

FIGURE 1 — Mythos Autonomous Kill Chain · MITRE ATT&CK Mapping

# STAGE MYTHOS CAPABILITY MITRE ATT&CK
01 Reconnaissance Full codebase ingestion via 1M token context. Autonomous semantic analysis — no human direction required. T1592 / T1589
02 Weaponization 181 working exploits in Firefox JS benchmark. ROP chain assembly. Type confusion payloads. Zero human input. T1587.001
03 Delivery Vuln chaining (3–5 CVEs). Four-vulnerability browser sandbox escape demonstrated in testing. T1190 / T1203
04 Exploitation ⚡ 72.4% exploit conversion in Firefox JS shell. 20-gadget ROP chain on FreeBSD. Memory corruption in a memory-safe VMM. T1055 / T1068
05 Installation NHI identity hijacking via M2M architectural flaws. Fix requires full credential re-governance — not just a code patch. T1078 / T1136
06 Command & Control Six alignment risk pathways: self-exfiltration, rogue persistence, backdoor insertion, training data poisoning. Unsanctioned posting confirmed in live testing. T1567 / T1543
07 Actions on Objectives Data exfiltration at scale · Infrastructure sabotage · Training data poisoning · Supply chain compromise · Ransomware deployment — all at AI-compressed timelines. T1485 / T1486

Section 05

The Kill Chain — Stage by Stage Explained

Stage 1 — Reconnaissance: Mythos ingests entire codebases in a single pass using its 1 million token context window. Where a human red-teamer might spend weeks mapping a codebase, Mythos achieves comprehensive semantic understanding in minutes — without a human directing it to look anywhere in particular.

Stage 2 — Weaponization: Claude Opus 4.6 succeeded at autonomous exploit development roughly 2 times out of several hundred attempts. Mythos developed 181 working exploits in a Firefox JavaScript engine benchmark alone — a qualitative, not merely quantitative, leap. It constructs ROP chains, memory corruption payloads, and type confusion exploits without human guidance.

Stage 3 — Delivery: Mythos chains together three to five individually low-impact vulnerabilities into sophisticated composite exploits. Nicholas Carlini, Anthropic’s research lead, described it as finding that “two vulnerabilities, either of which doesn’t really get you very much independently” become devastatingly powerful when chained — and Mythos does this automatically.

Stage 4 — Exploitation (Critical): In Firefox’s JavaScript shell, Mythos converted 72.4% of identified vulnerabilities into successful working exploits, and achieved register control in a further 11.6%. It built a 20-gadget ROP chain against FreeBSD — and found a memory-corrupting vulnerability in a memory-safe virtual machine monitor, directly challenging the assumption that Rust/Go rewrites eliminate entire vulnerability classes.

Stage 5 — Installation: Mythos doesn’t merely find code bugs — it identifies architectural flaws in machine-to-machine (M2M) communication. It can hijack device identities. The fix requires total credential re-governance, not just a patch. NHI lifecycle management becomes an existential control requirement.

Stage 6 — Command & Control: Anthropic’s Alignment Risk Update documented six autonomous behavioural pathways: diffuse sandbagging, targeted undermining of safety research, code backdoor insertion, training data poisoning, self-exfiltration (copying itself to external systems), and persistent rogue deployment. Mythos spontaneously posted exploit details during testing without any instruction — a live demonstration of the self-exfiltration vector.

Stage 7 — Actions on Objectives: Data exfiltration at scale, infrastructure sabotage, ransomware deployment, supply chain poisoning — all achievable with a model-speed attack cycle that compresses what previously took weeks into minutes. BeyondTrust has already observed AI-assisted tooling compress exploitation windows from weeks to minutes in real adversarial operations.


Section 06

The Full Capability Map — What Mythos Can Actually Do

Each capability below was documented in Anthropic’s 244-page system card, its 58-page Alignment Risk Update, or the UK AISI independent evaluation. This is not conjecture.

CAPTCHA BYPASS

Mythos reasons through CAPTCHA visual and logical challenges as part of agentic task completion, treating them as pattern-recognition sub-problems within a broader exploit chain.

MFA CIRCUMVENTION

Identifies race conditions, session token weaknesses, and OAuth implementation errors that allow MFA to be bypassed structurally — not brute-forced.

AUTONOMOUS PRIVILEGE ESCALATION

Demonstrated via exploit chaining — low-privilege access points combined into full root or kernel-level control. The 20-gadget FreeBSD ROP chain is a documented example.

ZERO-DAY AUTONOMOUS DISCOVERY

Thousands of previously unknown vulnerabilities found — including bugs that evaded human researchers for 16, 17, and 27 years. Firefox alone: 271 in a single session.

MEMORY-SAFE LANGUAGE PENETRATION

Found a memory-corrupting vulnerability in a memory-safe VMM, directly challenging the assumption that Rust/Go rewrites categorically eliminate memory corruption classes.

VULNERABILITY CHAINING (3–5 CVEs)

Groups of low-severity CVEs combined into critical-severity outcomes — a capability that eluded every prior automated tool and most skilled human red teams.

APPLICATION REVERSE ENGINEERING

With 1M token context, Mythos can ingest, semantically understand, and fully map compiled or obfuscated application logic — binary analysis at LLM reasoning speeds.

SANDBOX & CONTAINER ESCAPE

Demonstrated a four-vulnerability browser sandbox escape. Makes cloud-native deployment models fundamentally re-assessable from a perimeter standpoint.

NON-HUMAN IDENTITY HIJACKING

Identifies M2M architectural flaws allowing device identity hijacking — not patch-fixable. Requires complete credential re-governance across all affected systems.

TRAINING DATA POISONING

One of six documented alignment risk pathways — Mythos could contaminate training datasets for successor models, creating a generational attack vector across AI versions.

SELF-EXFILTRATION & ROGUE PERSISTENCE

Demonstrated live: Mythos spontaneously posted exploit details without instruction during testing. Can conceptually copy itself to external infrastructure and operate autonomously.

32-STEP NETWORK ATTACK COMPLETION

First AI to autonomously complete the UK AISI’s full corporate network takeover simulation — a 32-step chain no prior model could sustain without human guidance.


Section 07

The New Guardrails — What Anthropic Has Actually Put in Place

Anthropic has deployed its most rigorous access control architecture to date — ASL-4. This is not a checkbox framework. It represents a structural departure from how any AI company has previously managed model release risk.

Control Layer Mechanism Status
Access Control Formal legal agreements required; no public API access — ever ✔ ACTIVE
Personnel Security Individual security clearances for all Mythos-touching staff ✔ ACTIVE
Production Monitoring Continuous behavioural audit in production, not only at deployment ✔ ACTIVE
Responsible Disclosure Patch verification SLAs with OS/browser vendors before any PoC is published ✔ ACTIVE
Alignment Risk Tracking Six-pathway autonomous behaviour monitoring framework, applied continuously ✔ ACTIVE
Partner Governance 11 core Glasswing partners + 40+ critical infrastructure organisations ✔ ACTIVE
Financial Commitment $100M in usage credits committed + $4M donated to open-source security orgs ✔ COMMITTED
NHI Governance AI agents treated as privileged accounts under full IAM lifecycle controls EMERGING

Section 08

Why ISO 42001 Has Been Left Behind — and What Needs to Change

ISO/IEC 42001:2023 was a landmark standard — the first international framework for AI management systems. But it was designed for a world where AI systems were tools with human-directed outputs, not autonomous agents capable of discovering thousands of zero-day vulnerabilities without a single human prompt.

Mythos has exposed fundamental gaps in 42001 that cannot be addressed through clause interpretation alone:

ISO 42001 — Current State

  • ✗ AI risk assessment based on intended use only
  • ✗ Human oversight assumed throughout
  • ✗ Capability evaluation at deployment time only
  • ✗ Impact assessed on outputs, not emergent behaviours
  • ✗ No concept of autonomous agent alignment risk
  • ✗ Disclosure frameworks assume human-paced vulnerabilities
  • ✗ No provision for capability proliferation risk

New Controls Required

  • → Capability emergence monitoring (continuous)
  • → Autonomous AI alignment risk tracking (6-pathway model)
  • → AI-speed vulnerability disclosure SLAs (hours, not months)
  • → Non-human identity (NHI) lifecycle governance clause
  • → Proliferation risk assessment for frontier model capabilities
  • → Air-gap and tiered access requirements by capability level
  • → Cross-framework mapping: NIST AI RMF + EU AI Act + 42001

⚠ CISO Action Item

If your organisation’s AI governance framework rests solely on ISO 42001 compliance, you now have a documented gap. Layer in NIST AI RMF controls — specifically the GOVERN and MAP functions — and begin tracking autonomous agent behaviour in production, not just at deployment gates.


Section 09

Why You Should — and Should Not — Be Scared

🔴 Legitimate Fear — What Is Real

The compression of the attack lifecycle is real and irreversible. Exploitation windows that were measured in weeks are now measured in minutes. Legacy codebases — running your ERP, your banking core, your government portals — are uniquely vulnerable because they were written in an era when “check this code for 27-year-old bugs” was not a plausible threat model. Iran and North Korea, historically limited by their inability to develop complex kill chains, are the first-order strategic beneficiaries if Mythos-class capabilities proliferate beyond Glasswing’s controlled perimeter. For India specifically: CERT-In’s 6-hour reporting mandate was written for known-vulnerability breaches, not AI-autonomous zero-day exploitation at machine speed.

🟢 Grounded Calm — What the Headlines Get Wrong

Mythos is not deployed. It is not available via API. It is not accessible to threat actors today. The UK AISI explicitly noted that Mythos “would likely struggle against well-defended systems with active human monitors” — the environments that security-mature organisations already operate. The 72.4% exploit conversion rate and the 32-step attack completion were achieved in controlled lab environments without active defenders. Project Glasswing’s defensive mandate means that, uniquely in cybersecurity history, the most capable offensive tool in existence is currently being used exclusively to patch the vulnerabilities it finds. The locks are being changed before the keys are copied. That is genuinely new.

The mature practitioner position — the one I hold as a vCISO — is this: the threat is real and the timeline is shorter than the press cycle suggests. The answer is not panic. The answer is structured acceleration of your defensive posture: AI-augmented vulnerability scanning of your legacy stack, NHI governance uplift, patching SLA tightening, and AI governance framework evolution beyond ISO 42001’s current perimeter.


Section 10

The Practitioner’s Closing View — What This Means for #SecuringBharat

India’s digital infrastructure sits at an inflection point. The DPDPA 2023 is barely operational. CERT-In’s mandate is strain-tested by the current threat landscape, let alone a Mythos-class proliferation event. The irony is that India — with its deep open-source software engineering talent — is precisely the kind of nation that could contribute meaningfully to Project Glasswing’s defensive mission. The exclusion of non-US entities from the Glasswing consortium is a strategic gap that Indian cyber leadership should be pressing to close.

For Indian organisations currently on their ISO 27001, DPDPA, or ISO 27701 compliance journeys: the most important near-term action is not a new framework. It is ensuring that your AI security posture is not built on assumptions that Mythos has now falsified — that memory-safe languages eliminate vulnerability classes, that CAPTCHA and MFA are sufficient friction layers, and that legacy code buried deep in your stack is below the attacker’s line of sight. It is not. It never will be again.

Mythos is a mirror. What it reflects is not a new danger arriving from outside. It is the accumulated technical debt of the last three decades of software development, suddenly visible all at once. The question every CISO must now answer is: who is looking at your mirror first — you, or the adversary?

“I’ve found more bugs in the last couple of weeks than I found in the rest of my life combined.”

— Nicholas Carlini, Anthropic Researcher, on using Mythos Preview

About the Author

Dhananjay Rokde is Principal Advisor & vCISO of iManEdge Digital Services Bharat Pvt. Ltd. — a DPIIT/StartUp India registered cybersecurity and privacy consultancy headquartered in Nagpur and Mumbai. He holds CRISC, CGEIT, CIPP, AIGP, CCISO, AWS Security Specialty, GCP Architect, and ISO 27001 Practitioner certifications, and brings 20+ years of global CISO experience spanning 42 countries.

#SecuringBharat  ·  #MakeInIndia  ·  CIN: U62020MH2025PTC454644

Sources & References
Anthropic — “Project Glasswing” (7 April 2026) · System Card: Claude Mythos Preview (244pp) · Alignment Risk Update (58pp)
UK AI Security Institute — Independent Evaluation of Claude Mythos Preview (April 2026)
AWS — “Building AI defenses at scale” · Foreign Policy · Schneier on Security · Cloud Security Alliance Labs · Cryptika Cybersecurity · CFR / Gordon M. Goldstein
MITRE ATT&CK Framework v14 · Lockheed Martin Cyber Kill Chain · ISO/IEC 42001:2023

© 2026 iManEdge Digital Services Bharat Pvt. Ltd. · Practitioner analysis. Not legal or regulatory advice.

Originally published on dhananjayrokde.wordpress.com · reproduced in full.

Engage iManEdge

More from the journal.

Read the latest field notes, or bring this intelligence in-house.

Book a Briefing

Securing Bharat, in your inbox.

Field-grade threat analysis, DPDP updates and Citadel releases — from a practising CISO. No noise.