AI Can Now Hack Like a Pro: GPT-6 Cyber, DeepMind's RSA Warning and India's New AI Security Rules
OpenAI's GPT-6 Cyber built 948 Linux exploits, DeepMind's AI found weaknesses in 1024-bit RSA, and India's RBI and CERT-In moved on AI security in the same week. What it means and what organisations should do now.
By Pavan Kumar Verma · · 8 min read

This week, AI crossed a line the security world has been dreading for years.
On 5 November 2026, OpenAI unveiled GPT-6 Cyber, a model trained specifically to hunt for software vulnerabilities. Working through the Linux kernel — the code running most of the world's servers, Android phones and cloud systems — it built working exploits for 948 vulnerabilities. OpenAI rates it "Critical" for cybersecurity under its own safety framework, its highest risk level, and is releasing it only to selected security teams.
A day later, Google DeepMind announced that its AI agents, working with leading cryptographers, had found two new weaknesses in legacy 1024-bit RSA, the encryption still protecting some older systems.
And in India, the same week, the RBI reportedly hauled banks into a meeting over AI-assisted cyber-attacks, while CERT-In issued India's first AI Security Auditing Guidelines.
Put together, the message is clear: AI can now find and exploit software flaws faster and more cheaply than most human teams. Every organisation's security plans need to catch up.
A note on perspective: my company builds software and AI solutions, including for banks and insurers. Everything below is advice I'd give our own teams.
What happened this week
| Date | Development |
|---|---|
| 4 Nov 2026 | CERT-In issues AI Security Auditing Guidelines v1.0, a four-tier risk framework for AI systems used in India |
| Early Nov 2026 | RBI reportedly meets banks and issues a confidential advisory on API security after a wave of attacks on lenders' payment systems; the Home Ministry's I4C issues a public advisory on AI-assisted attacks on banks, NBFCs and fintechs |
| 5 Nov 2026 | OpenAI unveils GPT-6 Cyber, rated "Critical" for cyber capability |
| 6 Nov 2026 | Google DeepMind reports AI-assisted discovery of two weaknesses in legacy 1024-bit RSA |
GPT-6 Cyber: what it can do
According to OpenAI's own figures (no independent testing has been published yet):
- Linux kernel: the model scanned around 5,000 security-relevant locations, confirmed 1,724 vulnerabilities, and built working privilege-escalation exploits for 948 of them — more than five times its predecessor's 166 under the same setup. OpenAI says it reported 304 critical findings to kernel maintainers, with suggested fixes.
- Firefox: it found 81 previously unknown severe vulnerabilities. Mozilla accepted the reports and has shipped fixes.
- Cost: the whole Linux exercise reportedly cost about US$1.2 million — roughly US$1,280 per working exploit. For comparison, top hacking competitions pay tens of thousands of dollars per finding.
That last number is the one that should worry every security leader. Finding and weaponising a serious bug used to take a skilled expert weeks. It's becoming a line item.
OpenAI is limiting access to vetted customers in its "Daybreak" security programme, with IP-based restrictions, checks that the model's actions match the user's permissions, and monitoring for misuse. That's responsible. But capabilities rarely stay exclusive for long: other labs, open-source models and well-funded attackers will catch up.
DeepMind and the encryption warning
DeepMind's work, carried out with cryptographers including Nadia Heninger and Emmanuel Thomé, used AI agents to do much of the exploratory work. Reports describe two results:
- A much more efficient way to forge digital signatures for 1024-bit RSA — reportedly around 1,600 times better than the previous best-known attack in certain scenarios.
- A cheaper way to factor many related 1024-bit keys at once.
The important caveat: 2048-bit RSA and modern encryption are not affected. But 1024-bit keys still linger in old certificates, legacy hardware, industrial systems and embedded devices. The bigger lesson is speed: AI helped experts make advances in weeks that might otherwise have taken years.
Why this matters: the defender's dilemma
These tools cut both ways.
The good news: defenders can use AI to find and fix vulnerabilities before attackers do. OpenAI's 304 reports to Linux maintainers and Mozilla's Firefox fixes are exactly that — thousands of holes closed.
The bad news: attackers get the same advantages. Expect:
- More vulnerabilities discovered, faster — and a shorter window between a flaw being disclosed and attackers exploiting it.
- Cheaper, more targeted attacks on systems that were previously "not worth the effort".
- Pressure on patching. If AI floods vendors with bug reports, organisations will face more patches, more often.
- Legacy systems becoming liabilities. Old software and old cryptography were protected partly by the cost of attacking them. That protection is disappearing.
What's happening in India
RBI: API security under the spotlight
According to reports by MediaNama and Head and Tale, the RBI met banks in early November after a series of cyber-attacks in which criminals abused the APIs that banks use to let NBFCs, fintechs and payment companies move money. Individual frauds reportedly ranged from ₹3 crore to ₹15 crore.
The RBI's confidential advisory reportedly asks banks to:
- Keep a full inventory of API clients;
- Get NBFCs using their APIs audited by CERT-In-empanelled auditors — before onboarding new ones, and by 3 February 2027 for existing ones — at least once a year and after any major incident;
- Apply transaction limits immediately and velocity limits within a month;
- Use maker-checker or out-of-band confirmation for API debits and whitelist beneficiaries;
- Report incidents immediately.
(The advisory hasn't been published, so details may change.)
CERT-In: India's first AI security audit framework
CERT-In's AI Security Auditing Guidelines (version 1.0, 4 November 2026) apply to AI systems developed, procured, deployed or operated in India, or that process Indian data. Key points:
- Four risk tiers — Low, Moderate, High and Critical — based on a score across factors such as the sensitivity of data and the impact of decisions. MSMEs and startups are generally limited to the lower two tiers.
- Independent audits by CERT-In-empanelled auditors — at least annually for moderate-risk systems, more often for high-risk ones, and continuous monitoring for critical systems.
- Nine audit domains, from governance and data to model risk, bias, explainability, security, monitoring, human oversight and third-party suppliers.
- Specific scrutiny of AI agents: tool use, "excessive agency", least-privilege access, persistent memory and unsafe autonomous actions — plus prompt injection and data poisoning for generative AI.
- Deadlines to fix findings: for example, a remediation plan within 30 business days for moderate findings.
The guidelines don't spell out penalties, but they will quickly become the benchmark that regulators, auditors and customers use. And existing rules already apply: under CERT-In's 2022 Directions, cyber incidents — including those caused by an AI agent misusing credentials — must be reported within six hours, and logs kept for 180 days. CERT-In's Director General, Dr Sanjay Bahl, has said concern centres on "delegated authority without commensurate guardrails".
What organisations should do now
1. Patch faster
Assume the time between a vulnerability being disclosed and being exploited will shrink. Measure your patching time for critical systems in days, not months, and automate where you can.
2. Use AI for defence — carefully
Use AI tools to review code, scan for vulnerabilities and test your own systems before attackers do. Run them in isolated environments, with human review of findings and fixes.
3. Lock down your APIs
If you're a bank, NBFC, fintech or any business exposing APIs: inventory every client, enforce strong authentication, set transaction and rate limits, monitor for unusual patterns (such as ₹1 test transfers), and require extra confirmation for money movement.
4. Retire legacy cryptography
Find and replace 1024-bit RSA keys and other weak cryptography in certificates, devices and old systems. While you're at it, start planning for post-quantum cryptography — the same inventory work applies.
5. Classify your AI systems now
Map your AI systems against CERT-In's four tiers. Know which ones need independent audits, and budget for them before your customers or regulators ask.
6. Put guardrails around AI agents
Give agents task-specific, short-lived credentials; restrict their network access; require human confirmation for consequential actions; and keep a tamper-resistant log of every action. (More on securing AI agents here.)
7. Test your incident response
Can you detect, contain and report an incident to CERT-In within six hours? Rehearse it.
8. Check your suppliers
Your security is only as good as your vendors and partners — especially those connected to your systems through APIs. Ask for audit reports and contractual commitments on security and incident disclosure.
9. Invest in people
AI raises the bar for security teams. Train them to use AI tools, and hire for skills in application security, cloud security and AI risk.
What policymakers should do
- Turn guidelines into clear rules. CERT-In's framework is a strong start; sector regulators like the RBI, IRDAI and SEBI should align their requirements with it, so companies don't face conflicting demands.
- Build capacity for audits. India will need many more qualified AI security auditors, fast.
- Support open-source maintainers. Much of the world's infrastructure depends on volunteers maintaining code like the Linux kernel. If AI floods them with vulnerability reports, they'll need funding and help to fix them.
- Set rules for "dual-use" AI models. Governments should work with AI labs on who gets access to offensive-capable models, and how misuse is detected and reported.
Final thought
For decades, cybersecurity has been a contest between human attackers and human defenders. That era is ending. AI now finds flaws in hours that experts took weeks to find, and builds exploits for the price of a laptop.
That's a threat — and an opportunity. Organisations that use AI to find and fix their own weaknesses first, patch faster, retire legacy systems and put guardrails around their own AI will be safer than ever. Those that don't will find that the cost of attacking them has just fallen dramatically.
How is your organisation preparing for AI-powered attacks? Share your experience in the comments.
Sources: The Decoder, "OpenAI's GPT-6 Cyber can both break into systems and propose how to fix them" (5 November 2026); Cybersecurity News, GBHackers and Technology.org on GPT-6 Cyber (November 2026); Crypto Briefing, "Google DeepMind's AI agents expose two weaknesses in legacy 1024-bit RSA" (6 November 2026), Pushmeet Kohli (Google DeepMind) on X, and Byteiota on the RSA findings; MediaNama, "RBI meets bankers amidst rising AI cyber-attacks, mandates security audits" (6 November 2026), citing Head and Tale; Business Standard on RBI's scrutiny of banks' cyber defences (October 2026); CERT-In, AI Security Auditing Guidelines v1.0 (4 November 2026); MediaNama's lowdown on the guidelines and interview with CERT-In Director General Dr Sanjay Bahl (November 2026); SANSA Legal on the guidelines' four-tier risk model; CERT-In Directions under Section 70B(6) of the IT Act (28 April 2022).