Tuesday, November 17, 2026 AI news, turned into opportunities PDF newsletter
AI Opportunity Daily
Subscribe
CryptographyAI musicCloud GPUs

AI finds cracks in old RSA keys, Elvis's owners sue Suno, AWS splits GPUs seven ways

An AI lab used agents to find new weaknesses in an old encryption standard, music owners opened a fresh legal front against an AI song generator, and AWS made it cheaper to run several small models on one GPU. The thread: AI is speeding up work on every side, so old keys, missing permissions and idle hardware all cost more than they used to.

Key takeaways

  • Google DeepMind says AI agents found two theoretical weaknesses in legacy 1024-bit RSA; 2048-bit keys are fine, but old keys are now a clearer liability.
  • Elvis Presley's rights owners and Eminem's publisher sued Suno, making voice cloning, not only training data, a front line in AI music law.
  • AWS SageMaker can now split one GPU into up to seven slices, which can cut serving costs for small models when traffic is steady.
Story 1 of 3Research5 min read

Google DeepMind says AI agents found two new weaknesses in legacy 1024-bit RSA

DeepMind said on 6 November that AI agents, steered by human cryptographers, produced a theoretical 1,600x improvement in forging some 1024-bit RSA signatures and a 30x cheaper way to factor batches of keys. It stresses that 2048-bit keys are not affected.

Google DeepMind announced on 6 November that AI agents working under human supervision had produced two advances against 1024-bit RSA, an older key size of the encryption that has protected web traffic, software updates and logins for decades. The announcement came from Pushmeet Kohli, who leads DeepMind's science work, in posts on X, and was reported by Crypto Briefing, OfficeChai and others. DeepMind researchers Dj Dvijotham and Abhradeep Thakurta led the work with two well-known academic cryptographers, Nadia Heninger and Emmanuel Thomé. DeepMind says the advances took weeks rather than the years such work normally takes.

The first result concerns signature forgery: faking the digital seal that proves a message or file came from the owner of a key. By DeepMind's own theoretical estimates, the new method is roughly 1,600 times more efficient than the best previous technique for 1024-bit keys, in cases where an attacker controls only a small part of the signed data, such as a 32-byte fingerprint. The second is a batch-factoring algorithm that cuts the average cost of breaking each additional 1024-bit key in a large batch by about 30 times. The one result described as tested: the agents reproduced an earlier academic attack on an already-solved benchmark key using 3.5 times less CPU time.

The work lands in a busy year for attacks on old RSA. In September a paper co-authored by Heninger and Thomé (IACR ePrint 2026/2131) showed that 1024-bit RSA signatures can be forged without factoring the key, given temporary access to a raw signing service, at a reported cost of roughly 1,380 CPU core-years. Other AI companies have also publicised AI-assisted factoring records on smaller RSA numbers this year. Standards bodies have discouraged 1024-bit RSA for more than a decade, yet it survives in old routers and VPN boxes, firmware signing, internal certificates, unrotated SSH keys and hardware security modules set up years ago.

The limits matter. Kohli said the results only affect obsolete 1024-bit systems, and that modern protocols, which almost universally use 2048-bit or larger keys, remain secure. The 1,600x figure is a theoretical estimate, not a demonstrated attack, and nobody has reported forging a real certificate with it. The batch saving applies to factoring many keys together, not to breaking a single key cheaply. At the time of writing we could not find a full DeepMind technical paper; coverage relies on Kohli's posts and secondary reports, so the specific numbers should be treated as DeepMind's own claims until independent cryptographers publish a review.

The bigger signal is speed. If AI agents can compress years of cryptanalysis into weeks, the gap between theoretically weak and practically broken shrinks for every old system still in service. Kohli argued that the mathematical hardness of encryption schemes should no longer simply be assumed. For organisations, that is a practical prompt to find the 1024-bit keys still hiding in their systems, replace them, and start planning the larger move to post-quantum cryptography. For freelancers and small security firms it creates steady, unglamorous work: inventories, key rotations and written evidence that the job has been done.

Why it mattersAI is shortening the time it takes to turn theoretical weaknesses in encryption into practical ones, so old keys that 'probably still work' are becoming a real liability.

👀 What to watch

Whether DeepMind publishes full papers, how independent cryptographers rate the 1,600x estimate, and whether vendors and regulators set firmer deadlines to remove remaining 1024-bit keys.

3 opportunities from this story

1

Legacy-key audits for small businesses, schools and SACCOs

Low⏱ 1–2 weeks💰 Fixed fee per audit, plus paid remediation work

Offer a fixed-price check that finds 1024-bit RSA keys and weak certificates across a client's websites, email servers, VPNs, routers and SSH servers, then delivers a one-page remediation plan. Many small organisations run old appliances and have no security staff, so nobody has looked. A Google headline gives you an easy, honest conversation opener, and the fixes are well understood.

Best for
Freelance sysadmins, IT support firms and junior security consultants
First step this week
Scan your own servers (or a friendly client's, with written permission) and turn the results into a sample one-page report.
Open the full playbook
Launch steps
  1. Build a checklist: TLS certificates, SSH host and user keys, VPN and firewall settings, code-signing keys, DNSSEC
  2. Script the scans with open-source tools so each audit takes hours, not days
  3. Create a plain-English report template with a red/amber/green summary
  4. Offer key replacement and certificate renewal as a paid add-on
  5. Pitch 20 local businesses with a free external check of their public website
Tools
OpenSSLtestssl.shNmapssh-auditssh-keygen
Risks

Only scan systems you have written permission to test; scanning others can break the law. Most modern websites will already be fine, so sell peace of mind and a clean record rather than promising to find disasters, and do not overstate the threat: DeepMind says 2048-bit keys are safe.

2

Managed certificate and key rotation for IT providers

Medium⏱ 3–6 weeks💰 Monthly retainer per client, priced by number of systems

Turn one-off audits into a monthly service: keep an inventory of every certificate and key a client uses, send expiry alerts, rotate weak keys on a schedule and maintain a short roadmap toward post-quantum cryptography. Expired or weak certificates are a common cause of outages and failed security reviews. Recurring revenue suits managed service providers that already look after a client's machines.

Best for
Managed service providers and IT support companies
First step this week
Build a certificate-and-key inventory spreadsheet for one existing client and show them what is expiring or weak in the next 12 months.
Open the full playbook
Launch steps
  1. Standardise an inventory format (system, key type, size, owner, expiry)
  2. Automate renewals where possible with ACME certificates
  3. Write rotation runbooks with rollback steps for each system type
  4. Send clients a monthly one-page status report
  5. Add a yearly crypto-readiness review as an upsell
Tools
certbot (Let's Encrypt)Smallstep step-caHashiCorp VaultGoogle Sheets or Airtable
Risks

A botched key rotation can take a client's systems offline. Use staged changes, maintenance windows and tested rollback plans, and set out responsibility clearly in the contract.

3

Plain-English workshops on AI-speed cryptanalysis for IT managers

Low⏱ 1–2 weeks💰 Paid workshops, corporate training fees and audit leads

Run a paid half-day workshop or short online course that explains what 1024-bit versus 2048-bit means, how to check your own systems in an afternoon, and how to plan for post-quantum encryption. Most coverage of stories like this is either too technical or too alarming. A calm, practical session is useful to banks, SACCOs, hospitals and universities, and it doubles as lead generation for audit work.

Best for
Security trainers, tech writers and IT consultants
First step this week
Draft a 10-slide outline and a one-page checklist, then offer a free 45-minute webinar to test interest.
Open the full playbook
Launch steps
  1. Write a short explainer with sources and a self-check checklist
  2. Record a 10-minute demo of checking key sizes
  3. Host a free webinar and collect sign-ups
  4. Package a paid in-house version for organisations
  5. Update the material as DeepMind and others publish papers
Tools
Google Slides or CanvaZoom or Google MeetLoomSubstack
Risks

It is easy to slide into scaremongering, which damages trust. Stick to the facts (2048-bit keys are safe), cite sources and narrow your niche, for example East African financial institutions, because general security content is crowded.

Story 2 of 3Policy4 min read

Elvis Presley's rights owners and Eminem's publisher sue Suno over AI voice clones and training

Two federal lawsuits filed on 6 November accuse AI music generator Suno of training on copyrighted songs and letting users clone Elvis Presley's voice; Eminem's publisher seeks about $41 million over 269 songs. Suno denies wrongdoing.

Two new lawsuits filed on Friday 6 November target Suno, one of the most popular AI music generators. A group of Elvis Presley rights owners, including Authentic Brands Group, whose unit Elvis Presley Enterprises runs the Presley business, and Raleigh Music Publishing, sued in federal court in the Northern District of California, Billboard and Variety report. Separately, Eight Mile Style, the publisher that controls many of Eminem's songs, filed its own copyright suit. The same legal team, which includes Richard Busch, represents both sets of plaintiffs. Eminem himself is not a party to the case.

The Presley complaint, reported to run to 51 pages, goes beyond copyright. It argues that Suno lets users recreate Presley's distinctive voice and claims that, as recently as last month, the service produced an Elvis-style vocal from lyrics that simply included his name. It also invokes Tennessee's ELVIS Act, a 2024 state law that protects a performer's voice from unauthorised AI imitation. Eight Mile Style's suit focuses on training: it says Suno copied its catalogue without permission and claims $150,000 for each of 269 compositions, about $40.4 million, with reports putting the total above $41 million once other claims are added.

Suno already faces a long list of legal fights. Universal Music Group and Sony Music have sued it, as has a class of independent musicians led by Jason Isbell, according to Variety. Warner Music Group took a different path and settled with Suno through a licensing deal in 2025. The Presley rights owners also sued Anthropic last month, claiming its Claude chatbot was trained on 274 Presley songs, and Eight Mile Style sued Meta for about $110 million in June, Billboard reports. Music owners are clearly testing every legal route at once, against both music specialists and general AI companies.

These are complaints, not rulings. Suno denies the claims; a spokesperson said the company exists to help people make new, original music rather than trade on anyone's name or likeness, and that it blocks prompts naming specific artists and songs. The plaintiffs say those filters are easy to get around. US courts have not settled whether training an AI model on copyrighted music is fair use, and the ELVIS Act has had little testing in court. The damages figures are what the plaintiffs are asking for, and similar disputes have often ended in licensing settlements rather than full payouts.

For anyone making or selling AI-generated music, voiceovers or jingles, the direction of travel is clear: a person's voice is becoming a separate, enforceable right alongside copyright, and US states are writing it into law. That raises the value of services that prove where audio came from and that the right permissions exist, and it makes 'sounds like a famous artist' a risky selling point. Creators and agencies using AI audio commercially should keep records of the tools, prompts and licences behind each track, and should avoid imitating any identifiable voice without written consent.

Why it mattersVoice cloning is becoming its own legal risk, separate from copyright, so anyone selling AI audio now needs permissions and a paper trail.

👀 What to watch

Suno's formal response in court, any fair-use rulings in the UMG and Sony cases, and whether more US states copy Tennessee's ELVIS Act.

3 opportunities from this story

1

Consent-based AI jingles and voiceovers with local singers

Medium⏱ 2–4 weeks💰 Fee per jingle or voiceover, with a fixed revenue share paid to the voice artist

Produce radio and social-media jingles for local brands using AI music tools, but with the voices of real local singers who sign a written licence and earn a share of every job. Brands get fast, affordable audio without the legal risk of sound-alike voices. Singers get a new income stream, and you get a clear selling point as voice rights tighten.

Best for
Music producers, studios and creative agencies
First step this week
Sign one local singer to a simple written voice-licence agreement and produce three sample jingles for an imaginary brand.
Open the full playbook
Launch steps
  1. Draft a voice-licence agreement covering use, duration, payment and the right to withdraw
  2. Record clean voice samples and build a consented voice model
  3. Produce a demo reel of three to five jingles in different styles
  4. Pitch radio stations, ad agencies and SMEs that advertise on WhatsApp and TikTok
  5. Keep a record of the tools, prompts and licences behind every delivered track
Tools
ElevenLabs (consented voice cloning)Reaper or FL StudioA licensed AI music toolDocuSign or a signed PDF contract
Risks

Tool terms and laws can change; get a lawyer to review the licence and be open with clients that AI is used. Some buyers will still choose cheaper knock-offs, so compete on quality, speed and legal safety rather than price.

2

AI-audio rights checks for agencies and YouTubers

Low⏱ 1–2 weeks💰 Fixed-fee audit, plus a paid template and policy pack

Offer a fixed-fee review of the AI-generated music and voice in a client's content: which tools were used, what their licences allow, whether any track imitates a known artist, and what records are missing. Then deliver a simple policy and a logging template. Agencies and creators with large back catalogues are exposed and rarely have anyone checking.

Best for
Content strategists, paralegals and creative operations freelancers
First step this week
Write a two-page checklist covering the main AI music tools' commercial terms and test it on your own or a friend's channel.
Open the full playbook
Launch steps
  1. Summarise the commercial-use terms of the main AI music and voice tools
  2. Build an audit spreadsheet: asset, tool, plan, licence, consent, risk level
  3. Create a one-page AI audio policy template
  4. Offer the audit to five agencies at an introductory price
  5. Partner with a media lawyer for cases that need legal advice
Tools
Google Sheets or AirtableNotionC2PA Content Credentials toolsThe tools' published terms of service
Risks

This is not legal advice, and you should say so; refer serious cases to a lawyer. Rules differ between Kenya, the US and the EU, so state clearly which ones your checklist covers.

3

Rights-cleared training-data catalogues from local musicians

High⏱ 2–4 months💰 Licence fees from AI companies, shared with rights holders after a platform cut

Gather recordings from East African producers and artists who own their rights, clean up the metadata and offer the catalogue for licensed AI training, with fair revenue splits. Lawsuits like these push AI music companies toward licensed data, and catalogues from under-represented regions and genres are scarce. This is a slower, harder business, but it puts musicians on the paying side of AI.

Best for
Music publishers, collecting-society insiders and music-tech founders
First step this week
Interview ten local producers about whether they would license their catalogues for AI training and on what terms.
Open the full playbook
Launch steps
  1. Design a clear opt-in licence with payment terms and an exit clause
  2. Verify ownership for every track, including co-writers and labels
  3. Standardise metadata (genre, tempo, language, instruments, stems)
  4. Build a small sample set to show buyers
  5. Approach AI music companies and research labs that say they use licensed data
Tools
A rights-management spreadsheet or databaseAudio tagging toolsCloud storageE-signature software
Risks

Buyers are few and deal sizes are uncertain, and clearing rights is slow and complex. A single disputed track can sink trust, so verify ownership carefully and pay artists transparently.

Story 3 of 3Tools4 min read

AWS lets one GPU serve up to seven AI models at once on SageMaker, and publishes cost numbers

Amazon SageMaker real-time endpoints can now split a single Nvidia GPU into as many as seven isolated slices. In AWS's own tests, an H100 slice beat a whole A100 for an 8-billion-parameter model and cut cost per million output tokens by 18–34% at busier loads.

Amazon Web Services has added GPU partitioning to SageMaker AI real-time inference endpoints, its managed service for running AI models in production. The feature uses Nvidia's Multi-Instance GPU (MIG) technology to carve one physical GPU into up to seven smaller, isolated slices, each with its own compute and memory, so several models can share a card without slowing each other down. The feature was announced at the end of October, and on 6 November AWS published a detailed technical blog post with setup instructions and benchmark results, which Unite.AI and others covered.

Support depends on the hardware. P4d instances (Nvidia A100 40GB) and P5 instances (H100 80GB) split into up to seven partitions per GPU; the newer G7e (RTX PRO 6000 Blackwell) splits into four and G7 into two. On an eight-GPU p4d.24xlarge, AWS shows a layout that yields 24 partitions in total. In AWS's own testing with Meta's Llama-3.1-8B-Instruct, a single H100 partition produced 53–89% more output tokens per second than a whole A100-40GB, and cost 18–34% less per million output tokens once traffic passed a break-even point. Carving up the GPUs added roughly 17 seconds to provisioning.

The problem it tackles is common. Many useful models, such as small language models, embedding models, rerankers and speech models, use only a fraction of a large GPU, but customers pay for the whole card by the hour. Teams have worked around this by packing models into one container or using Kubernetes tools, which can let one busy model starve the others. MIG has existed on Nvidia's data-centre GPUs since the A100, and AWS already offered GPU partitioning on its SageMaker HyperPod clusters. What is new is that it now works on the simpler managed endpoints that many smaller teams use.

There are real limits. Partitioning works only on endpoints built with inference components; single-model endpoints, serverless endpoints, asynchronous inference and mixed instance pools are excluded. The partition layout is fixed when the endpoint is created, so changing it means building a new endpoint and moving traffic. Isolation covers GPU compute and memory, while CPU, networking and storage are still shared. You still pay the full hourly price of the instance, and at low traffic a whole A100 was cheaper in AWS's tests. All benchmark figures are AWS's own and cover a single model size.

For builders, this nudges the economics of self-hosting. A small agency running a chat model, an embedding model and a speech model for several clients could put them on one GPU instance instead of three, if traffic is steady enough. That makes 'we host your own private model' offers more competitive with per-token APIs for some workloads, and it rewards people who can measure real traffic and pick the right slice sizes. The feature is available in seven AWS regions, including Ireland, Frankfurt and Singapore, so teams in East Africa would most likely serve from Europe.

Why it mattersPaying for a whole GPU to serve a small model is one of the main costs of self-hosting AI; slicing the card up makes small, multi-model deployments cheaper for steady workloads.

👀 What to watch

Whether Google Cloud and Azure offer the same on their managed endpoints, independent benchmarks beyond AWS's own, and further inference news at AWS re:Invent (30 Nov–4 Dec).

3 opportunities from this story

1

Private model hosting for several small clients on one GPU

Medium⏱ 3–6 weeks💰 Monthly hosting fee per client plus a one-off setup fee

Host small open models (a chat model, an embedding model for search, a transcription model) for several clients on one partitioned SageMaker instance, and sell each client a monthly plan. Clients get a private model that keeps their data off shared public APIs, and you spread the GPU cost across them. Partitioning means one client's traffic spike no longer slows down everyone else.

Best for
AI agencies, developers and small cloud consultancies
First step this week
Deploy one small open model on a SageMaker endpoint with inference components, measure its real cost per day, then delete it.
Open the full playbook
Launch steps
  1. Pick two or three small open models that cover common client needs
  2. Test partition sizes against realistic traffic for each model
  3. Price plans so that three or more clients cover the instance cost
  4. Write a data-processing agreement for clients
  5. Set up monitoring and billing alerts before onboarding anyone
Tools
Amazon SageMaker AIvLLM or Hugging Face TGI containersAmazon CloudWatchTerraform
Risks

GPU instances are expensive, so the model only works with several paying clients and steady traffic; start small and set billing alarms. You also take on data-protection duties, including under Kenya's Data Protection Act, so contracts and access controls must be in place.

2

GPU cost audits for AI startups

Medium⏱ 2–3 weeks💰 Fixed-fee audit, or a fee linked to verified savings

Review a startup's AI serving bill, find GPUs that sit mostly idle, and test whether partitioning or smaller instances would cut costs, then deliver a short report with measured numbers. Many young AI teams overprovision because they never had time to tune. AWS's own data shows savings only above a break-even load, so careful measurement is the product.

Best for
Cloud engineers, MLOps freelancers and DevOps consultants
First step this week
Reproduce a small version of AWS's benchmark on one model and publish your findings as a short case study.
Open the full playbook
Launch steps
  1. Get read-only billing and monitoring access from the client
  2. Map each model to its GPU usage and traffic pattern
  3. Benchmark candidate partition layouts with realistic load
  4. Write a report with options, expected savings and migration steps
  5. Offer the migration itself as a follow-on project
Tools
AWS Cost Explorergenai-benchAmazon SageMaker AIPython and Jupyter
Risks

Savings depend on traffic and may not appear at low volumes, so never promise a number before measuring. Ask only for the access you need and follow the client's security rules.

3

SageMaker GPU-partitioning starter kits and tutorials

Low⏱ 1–2 weeks💰 Paid templates, a short course and consulting leads

Publish a clear tutorial and sell a ready-made template that sets up a partitioned endpoint with two or three models, monitoring and cost alarms. AWS's documentation is detailed but long, and many developers want a working example they can copy. The kit also builds credibility and brings in consulting leads.

Best for
Developer educators, technical writers and cloud freelancers
First step this week
Build a working two-model partitioned endpoint, record a 10-minute walkthrough and post it free.
Open the full playbook
Launch steps
  1. Build and test the template in one region
  2. Write a tutorial covering setup, limits and clean-up
  3. Publish a free version on GitHub
  4. Sell an extended kit with monitoring and cost alarms
  5. Update it when AWS changes the feature
Tools
AWS CDK or TerraformGitHubYouTube or LoomGumroad
Risks

AWS may publish its own templates, and the feature may change. Running benchmarks costs money, so keep tests short and always delete endpoints afterwards.

Get these as a PDF every 3 days

Free. One email every 3 days. Unsubscribe any time.

More editions