The models: What has actually changed

Claude Mythos Preview and GPT-5.5 Cyber (referred to herein as GPT-5.5 Cyber or Daybreak) are not enhanced signature engines or smarter Security Orchestration, Automation and Response (SOAR) playbooks. They are autonomous reasoning systems capable of end-to-end offensive and defensive cyber operations — and their capabilities have been independently verified at scale.

Anthropic launched Project Glasswing in April 2026, initially restricting Mythos Preview to approximately 50 partner organisations. The rationale was unambiguous: the model could autonomously discover and exploit software vulnerabilities without human guidance, and releasing it publicly risked placing a supercharged cyberweapon in adversarial hands [1]. Within four weeks, those 50 partners had collectively surfaced over ten thousand high- or critical-severity vulnerabilities across the most systemically important software in the world [1].

The numbers are operationally staggering. Cloudflare found 2,000 bugs — 400 of them high- or critical-severity — with a false-positive rate that its team rated better than human testers [1]. Mozilla fixed 271 vulnerabilities in Firefox 150, more than ten times the number found in the previous release [1]. The UK’s AI Security Institute confirmed Mythos Preview is the first model to autonomously complete both of its cyber ranges — multistep simulations of full corporate network attacks — end to end [1].

OpenAI’s response — GPT-5.5 Cyber, announced on 7 May 2026 — takes a different structural approach [6]. Rather than a single restricted model, OpenAI created a Trusted Access for Cyber (TAC) framework: a tiered identity-and-trust system that gates capability by the verified status of the user. Standard GPT-5.5 handles general defensive tasks. GPT-5.5 with TAC unlocks vulnerability triage, malware analysis, binary reverse engineering, and patch validation for verified defenders. GPT-5.5 Cyber — available in limited preview to critical infrastructure operators — adds authorised red-teaming and live exploit validation [6]. Phishing-resistant authentication is mandatory for the most permissive tier [6].

What both programmes share is a foundational claim described by LSE researchers Dr Beatriz Lopes Buarque and Microsoft AI engineer Abdullah Abu-Hassan as a collapse in the cost of offensive cybersecurity [5]. Capabilities that once required an elite operator with decades of experience can now be executed with a few lines of code. An elite cyber capability has become, in effect, a software function. The Mythos system card confirms the model can run end-to-end attacks on enterprise networks with weak security postures, and complete a 32-step corporate network attack simulation in timeframes that would take a skilled human twenty hours [5].

India’s position: Inside the perimeter, behind the curve

India is part of Project Glasswing. That much is confirmed. What is less discussed — and more consequential for enterprise CISOs — is the nature of that participation and what it reveals about India’s structural position in the frontier AI security landscape.

The access lag

When Anthropic launched Project Glasswing in April 2026, the founding cohort of approximately 50 organisations was drawn almost entirely from the United States and United Kingdom [7]. Amazon, Google, Microsoft, Apple, NVIDIA, CrowdStrike, Palo Alto Networks, and the UK’s AI Security Institute were scanning and hardening critical software with Mythos Preview while the rest of the world waited. India was not at that table.

During those weeks, Indian organisations that recognised the risk were conducting controlled security tests using Claude Opus 4.7 — Anthropic’s preceding flagship model — as a substitute [4]. Major IT services firms including Infosys and Tata Consultancy Services ran these tests. CERT-In reviewed critical systems including Aadhaar and government login platforms using the same [4]. That is the access gap: India’s most important cybersecurity institutions spent the most critical weeks of the frontier AI security transition working with a less capable tool, scanning the same infrastructure that a Mythos-class adversary could target.

By late May, Anthropic expanded the programme to approximately 150 organisations across more than 15 countries [7]. India was included in this second wave. According to reporting by The Indian Express, the Indian organisations now understood to have received access to Mythos Preview include: the Indian Cyber Crime Coordination Centre (I4C); CERT-In; the National Critical Information Infrastructure Protection Centre (NCIIPC); and the Department of Telecommunications’ Digital Intelligence Platform (DIP) [7]. NCIIPC — which reports to the National Security Advisor in the Prime Minister’s Office — and CERT-In had specifically requested access to scan banking and power infrastructure [7]. Discussions are also underway to extend access to cybersecurity research institutions and dedicated AI security teams within India’s largest IT services companies [7].

Inclusion in the second wave is meaningful. It is not equivalent to founding membership. The organisations that joined in April have six to eight weeks of operational experience with Mythos Preview — six to eight weeks of vulnerability findings surfaced, triaged, and patched — that Indian institutions do not. Every day of that lag is a day during which vulnerabilities in systems that underpin Indian enterprise environments sat unexamined by the most capable scanning tool ever deployed. Some will already be in Anthropic’s coordinated disclosure queue. Others will not surface for months.

The sovereign hosting constraint

India’s access position is further complicated by the government’s push for sovereign hosting of Claude AI — specifically for use in banking, telecom, and critical infrastructure, where jurisdictional and national security concerns around foreign-hosted AI infrastructure are acute [4]. This ambition is legitimate and directionally correct. Its near-term consequence is a meaningful delay in the timeline for regulated sector organisations to deploy Mythos-class capability at production scale within compliant data boundaries.

For CISOs in banking, insurance, and telecom — the sectors where Mythos-related regulatory guidance from SEBI and RBI is already emerging — the sovereign hosting question is not an abstract policy debate. It is the difference between being able to deploy AI-assisted vulnerability scanning on production systems and being limited to air-gapped pilots on non-production environments. Until the sovereign hosting framework is resolved, regulated CISOs face a compliance constraint that their global peers do not.

The adversary does not have the same constraint

The access lag and the sovereign hosting delay apply to Indian defenders. They do not apply to adversaries.

India’s threat landscape is specific and well-documented. State-sponsored groups operating from China and Pakistan have demonstrated sustained interest in Indian critical infrastructure — power grids, banking systems, government networks, and telecom. Ransomware operators have increasingly targeted Indian financial institutions and IT service providers, which collectively manage infrastructure for much of the region. Supply-chain attackers targeting Indian IT services firms gain downstream access to the clients those firms manage.

Against these adversaries, the LSE researchers’ warning carries particular weight: the Mythos system card confirms a plausible risk that the model could carry out actions leading to substantially higher odds of a global catastrophe [5]. More immediately, it is the first model to complete a 32-step corporate network attack simulation autonomously — the kind of multistep operation that previously required a coordinated team. When equivalent capability reaches adversarial hands — through state development, through leaked weights, through a less safety-conscious competitor releasing an open model — the attack surface facing Indian enterprises will have changed in kind, not just degree.

Anthropic is candid on this timeline: models as capable as Mythos Preview will soon be developed by many different AI companies [1]. The question for Indian CISOs is not whether that happens. It is whether Indian enterprises will have hardened their environments in time.

Implications for the CISO role: A dimension-by-dimension assessment

Threat intelligence and risk assessment

Your threat intelligence baseline has changed. Vulnerabilities that previously required months of skilled manual exploitation research can now be surfaced and weaponised in hours. CERT-In’s existing patch windows were calibrated for a world where exploit development was expensive and slow. That world no longer exists.

Mythos Preview demonstrated this concretely: it constructed a working exploit for a vulnerability that had existed in the OpenBSD operating system’s codebase — one of the world’s most security-conscious projects — for 27 years, undetected [5]. The vulnerability class is not exotic. OpenBSD underlies infrastructure across the Indian banking and government stack. When similar findings from Anthropic’s open-source scanning enter the 90-day coordinated disclosure window, they will become public. The patch race begins at that point, and it is a race between defenders who know the window and adversaries who will know it simultaneously.

The immediate action: review every patch SLA for critical and high-severity vulnerabilities. The window that was acceptable under the old threat model is not acceptable now. Any lag must be treated as a residual risk requiring explicit board sign-off and documented compensating controls.

Security operations: Detection and response

Adversaries with Mythos-class capability will not attack at human speed. They will deploy autonomous agents that simultaneously operate across multiple hosts and networks without human command [5]. The LSE researchers confirm that Mythos Preview is the first model to complete a 32-step corporate network attack simulation end-to-end — operations that would take a skilled human twenty hours, executed autonomously [5].

SOC workflows calibrated for human-speed attack progressions cannot detect or respond to this threat profile in time. An autonomous attacker can progress from initial access to lateral movement to data exfiltration before a human analyst has finished reading the first alert. Indian enterprise CISOs — whether running in-house SOCs, hybrid models, or outsourced managed security services — must assess whether their detection logic and response playbooks are designed for this reality.

The compensating controls are not new. They become more urgent. Harden default network configurations. Enforce multi-factor authentication across every privileged access pathway. Maintain comprehensive logs. Implement zero-trust segmentation. These controls buy time when a patch has not yet landed. In a Mythos-class threat environment, time is the only thing that matters [1].

The defender dividend: What AI does for the blue team

The picture is not only threat. The same capability that expands adversarial attack surface also gives defenders something genuinely unprecedented: the ability to find vulnerabilities in their own systems before adversaries do, at a scale and speed no human team could match.

The Glasswing data is compelling on its own terms. Partners report bug-finding rates increasing by more than a factor of ten [1]. At one Glasswing partner bank, Mythos Preview detected and stopped a $1.5 million fraudulent wire transfer after a threat actor compromised a customer email account [1]. Vendors embedded in the Glasswing and OpenAI TAC ecosystems — IBM, Qualys, CrowdStrike, Palo Alto Networks, SentinelOne — are translating these capabilities into enterprise security products that Indian organisations already buy [2][3][6].

The key analytical point from Qualys CTO Dilip Bachwani is worth retaining: the most significant impact of frontier AI may not be the threats it creates, but the speed at which it changes the economics of cyber risk [3]. Traditional vulnerability management was designed for human-speed discovery and human-speed remediation. Both sides of that equation have now changed. Defenders who access AI-assisted scanning gain an asymmetric advantage — but only if they have the human capacity to act on what the AI surfaces.

Vendor and supply chain risk

Frontier AI capability is not only arriving through direct Glasswing or TAC programme access. It is arriving embedded in the security products your vendors are already shipping. IBM Concert, Qualys Enterprise TruRisk Management, CrowdStrike, Palo Alto Networks, SentinelOne, Snyk — these platforms are integrating Mythos-class and GPT-5.5 capabilities as standard features [2][3][6].

This creates a dual obligation. First, understand what frontier AI capabilities are now embedded in your existing security stack, what data those capabilities process, where processing occurs, and what your vendor’s human-oversight model looks like. SEBI has explicitly flagged that AI-based vulnerability tools raise concerns around data confidentiality and output reliability [4] — that concern applies to capability embedded in your vendors’ platforms, not just tools you procure directly. Second, and more structurally: if your managed security service provider is deploying frontier AI scanning on your environment without a documented validation and escalation process, you are not gaining a defender advantage. You are generating findings that may not be acted on correctly.

Regulatory compliance and data governance

The Indian regulatory environment is moving at pace. SEBI’s cyber-suraksha.ai task force has issued guidance covering immediate patching requirements, AI-based vulnerability assessment processes, stronger vendor coordination, stricter API controls, enhanced monitoring, zero-trust frameworks, and mandatory onboarding onto market security platforms [4]. The RBI is coordinating with global regulators on Mythos-related risks in real time [4].

The Digital Personal Data Protection Act 2023 creates binding constraints on processing personal data with foreign-hosted AI. Any frontier AI security tool that ingests logs, network telemetry, source code, or configuration data from production systems containing personal data triggers DPDPA obligations. The government’s sovereign hosting ambition — still unresolved — signals that those constraints may tighten further for regulated sector applications [4].

CISOs in banking, insurance, capital markets, telecom, and healthcare should treat regulatory mapping as a prerequisite to procurement, not a post-implementation compliance exercise. The mapping must cover DPDPA 2023, sector-specific RBI and SEBI cybersecurity directions, and applicable MeitY guidelines. Engage your legal and compliance teams now. The regulatory framework is being written around decisions you are making this quarter.

Board reporting and cyber risk governance

Frontier AI creates a reporting obligation and a reporting opportunity simultaneously. The obligation: your board needs to understand that the threat environment has materially changed, that existing patch timelines and SOC playbooks may be calibrated against a threat model that is already obsolete, and that the regulatory landscape is shifting in real time. Finance Minister Sitharaman and IT Minister Vaishnaw flagged the Mythos risk publicly in April [7]. If the finance minister has flagged it, your board should already have been briefed.

The opportunity: CISOs who can frame the defensive case clearly — what AI-assisted scanning could surface in your environment, what closing those vulnerabilities is worth in reduced breach risk, what the cost of not acting is — will strengthen their position as strategic advisors rather than operational managers.

Frame the board conversation around three questions. What is our current exposure to vulnerabilities that a Mythos-class adversary could find and exploit? What is our plan and timeline to close the most critical gaps? And what is our path to accessing defensive AI capability proportionate to the threat we now face?

Talent, skills, and the AI security analyst

The skills required to deploy, validate, and govern frontier AI security outputs do not map onto existing job descriptions. The analyst this era requires is part vulnerability researcher, part model-output critic, part governance specialist. They must be able to recognise a plausible hallucination in a triage finding, contextualise AI output against a live organisational threat model, and document the human-in-the-loop decision trail that regulators will eventually require.

This profile is scarce everywhere. It is acutely scarce in India. The immediate risk for Indian enterprises dependent on managed security service providers is that those providers deploy frontier AI tools at volume without having built the internal expertise to validate what those tools produce — generating noise that looks like signal, and signal that gets lost in the noise.

The near-term investment priority is targeted upskilling: vulnerability classification, AI output validation, prompt engineering for security workflows, and human-in-the-loop governance design. This is not a training-course purchase. It is a capability-building programme, and it needs to start before the tools are deployed, not after.

The cost and skills reality: What deployment actually requires

Access to frontier AI cybersecurity capability is not a procurement decision. It is a systems integration challenge with cost and skills barriers that most Indian enterprises are not currently positioned to clear without deliberate and sustained investment.

The cost of running frontier AI at security scale

Frontier AI models operate on usage-credit economics, not annual licences. Anthropic provided $100 million in usage credits to its initial Glasswing cohort of approximately 50 organisations — roughly $2 million per partner for exploratory usage [5]. Production-grade agentic scanning of large codebases at continuous cadence costs significantly more. Token costs for multi-step security reasoning tasks are materially higher than for standard generative AI workloads, and the agentic scanning pipelines that produce the most valuable findings are the most compute-intensive.

The remediation cost compounds this further. Anthropic’s own data shows each high- or critical-severity vulnerability found by Mythos Preview takes an average of two weeks to patch [1]. AI finds faster than humans can fix. Enterprises that deploy frontier AI scanning without proportionate investment in remediation engineering capacity will generate finding backlogs they cannot clear — widening the disclosure gap rather than closing it.

The hallucination risk: Understanding the accuracy picture

Frontier AI models are not infallible. In security contexts, an error is not an inconvenience — it is an operational risk.

Anthropic’s open-source scanning data shows that 9.4% of vulnerabilities initially assessed by Mythos Preview as high- or critical-severity were false positives on independent review [1]. That figure deserves both sides of its framing. A 90.6% true-positive rate at high-severity tier is, by any historical benchmark for automated tooling, extraordinary — Cloudflare’s team rated it better than human testers [1]. But in a regime where the model surfaces thousands of high-severity findings, a 9.4% false-positive rate still represents hundreds of findings that consume analyst time and, if acted on incorrectly, divert remediation resources from real vulnerabilities. SEBI’s explicit concern about output reliability [4] is operationally grounded.

The deployment principle is non-negotiable: AI amplifies analyst capacity; it does not replace analyst judgment. Every finding requires human validation before any remediation action is taken. The skills to perform that validation correctly are the scarcest input in the entire pipeline.

Strategic recommendations for Indian enterprise CISOs

The following recommendations are sequenced by urgency. Some require board sponsorship; all require action.

Immediate (0–90 Days)

• Review every patch management SLA for critical and high-severity vulnerabilities. The window acceptable under the old threat model is not acceptable now. Treat any gap as a residual risk requiring board sign-off and documented compensating controls.
• Map your open-source and third-party dependency exposure against the Glasswing vulnerability class. Operating systems, web browsers, cryptographic libraries, and network infrastructure libraries are all in scope. Findings will enter the public domain as the 90-day coordinated disclosure window closes — you need to know your exposure before they do.
• Interrogate your top five security vendors on frontier AI integration. Which capabilities have they embedded? What data do those capabilities process? Where is processing performed? What is their hallucination mitigation and human-override model? If they cannot answer clearly, that is itself a risk finding.
• Complete a DPDPA and sector-regulatory mapping before any AI security procurement. For banking, insurance, capital markets, telecom, and healthcare CISOs, this mapping is a compliance prerequisite. Engage legal and compliance teams immediately. The regulatory framework is being written around decisions you are making this quarter.
• Brief your board now, if you have not already. Finance Minister Sitharaman flagged this risk publicly in April [7]. Your board should not be hearing about frontier AI cybersecurity risk from a newspaper.

Medium Term (90–180 Days)

• Pilot AI-assisted vulnerability scanning in a bounded, non-production environment. Use generally available frontier AI models — Claude Opus 4.7 or GPT-5.5 with Trusted Access for Cyber — with explicit human validation at every triage step. Document false-positive rates and remediation timelines before scaling. Establish baseline metrics before the programme expands.
• Invest in targeted analyst upskilling: vulnerability classification, AI output validation, prompt engineering for security workflows, and human-in-the-loop governance design. This capability must precede deployment, not follow it.
• Review SOC detection logic and incident response playbooks against machine-speed attack progressions. Identify the specific gaps where human-in-the-loop response timelines will fail against an autonomous adversary. Prioritise automated detection in those gaps.
• Engage CERT-In, SEBI, and RBI proactively on your organisation’s posture and approach to frontier AI security tools. Early engagement provides regulatory visibility and positions your organisation as a responsible actor in a framework that is still being designed.

Strategic (180 Days and Beyond)

• Develop a frontier AI security governance framework: procurement criteria, deployment standards, validation requirements, human-override protocols, audit trails, and explainability obligations. Design it for the regulatory environment you expect in 18 months, not the one that exists today.
• Engage actively with cyber-suraksha.ai and CERT-In coordination initiatives. India’s regulatory and incident-response infrastructure for frontier AI security is being architected now. CISOs who participate help shape it. Those who wait inherit whatever was built in their absence.
• Build for capability parity, not vendor dependence. The strategic goal is an organisational capability to evaluate, deploy, validate, and govern whatever frontier AI security tools are available — not a dependency on a single vendor whose access terms, pricing, or regulatory status may change. Vendor lock-in in this space carries strategic risk.

Conclusion: The interim period is the danger zone

The arrival of Claude Mythos and GPT-5.5 Cyber does not create a future problem. It creates a present one. The specific asymmetry at its core: the ability to find vulnerabilities has accelerated by an order of magnitude; the ability to patch them has not. Anthropic acknowledges the bottleneck directly — it is human capacity to triage, report, and deploy fixes [1]. That constraint is identical for Indian enterprises.

India’s position adds a second asymmetry. The founding Glasswing cohort spent April and May closing vulnerabilities in software that Indian enterprises depend on. India’s government agencies, armed now with access to Mythos Preview, will begin the same work — but weeks behind, under sovereign hosting constraints that complicate regulated-sector deployment, and without the established remediation pipelines that the founding cohort has had time to build.

The interim period — while vulnerabilities are being rapidly discovered and slowly patched, while defenders are building capability and adversaries are working to acquire equivalent access, while regulatory frameworks are still being written — is the period of highest risk. The decisions Indian enterprise CISOs make in the next ninety days will shape their organisations’ exposure posture for years.

The tools themselves are extraordinary. But tools do not govern themselves. The human judgment required to deploy them responsibly, validate their outputs rigorously, navigate the regulatory landscape taking shape around them, and translate findings into board-level decisions that get funded — that is the CISO’s work. In the frontier AI era, it has never been more consequential.

References and Sources