Shared irresponsibilities and the importance of product privacy: Apple vs Microsoft – Mark Batchelor, Vibhuti Sinha, Chris Simmons, Gerry Gebel, Ajay Gupta, Tarvinder Sembhi – ESW #365

A. Cyberhaven Raises $88 Million to Protect Enterprise Data in the AI Economy

$88M Series C led by Adams Street Partners. "Cyberhaven protects the intellectual property that traditional data loss prevention (DLP), insider risk, and data security posture management (DSPM) tools fail to identify and secure—data like source code, product designs, and customer records. With its pioneering data lineage technology and foundational AI model that understands not only content but also context, Cyberhaven is uniquely able to classify any sensitive information, understand when it is at risk, and take action to protect it."

Are we really post-DSPM already? Sheesh, infosec marketing moves fast. Don't tell Tenable (wink)

B. ThreatModeler Raises $60 Million from Invictus Growth Partners

$60M in an institutional round from Invictus Growth Partners. Invictus also backed Binary Defense's last round in 2022.

ThreatModeler's products don't appear to do any threat modeling, however? I'm confused.

C. Greylock Leads $36 Million Financing for Cybersecurity Startup Seven AI

$36M Seed Round led by Greylock. Founding team from Cybereason. Autonomous threat hunting using AI. Name refers to the "Seven Patterns of AI" as outlined by Cognilytica.

• Hypersonalization (not a typo?)
• Recognition
• Conversation & Human Interaction
• Predictive Analytics & Decisions
• Goal-Driven Systems
• Autonomous Systems
• Patterns & Anomalies

D. Cybersecurity startup SpyCloud secures $35m to combat account takeovers

$35M round led by CIBC Innovation Banking, the investment arm of the Canadian Imperial Bank of Commerce. "SpyCloud specialises in detecting leaked employee login credentials and protecting consumer accounts through its platform."

E. Passwordless security company HYPR raises $30M to bolster defenses against AI security threats - SiliconANGLE

$30M venture round (last two rounds were Series C in 2021 and 2022) led by Silver Lake Waterman.

F. YesWeHack Raises 26 Million Euros to Accelerate Its Growth and International Expansion

€26M Series C led by Wendel. French bug bounty company that's also well known in French-speaking Canada. Also happens to have Renaud Deraison, the creator of Nessus on the board of directors.

"500 customers across 40 countries"

G. Stacklet Secures $14.5M in Funding to Revolutionize Cloud Governance and Optimization, Slashing Enterprise Spend and Risk

$14.5M Series B led by SineWave Ventures, bringing total funding to $36.5M. Stacklet was co-founded by the team behind CNCF's (Cloud Native Computing Foundation) Cloud Custodian open source project and community.

The project isn't pure play security, but anything focused on observability and governance is going to have a strong security selling point, particularly with cloud. Arguably, most cloud issues are tied to observability and governance anyway.

We don't know the deal amount, but as a public company, I'd be surprised if we don't see it in an 8K or the next 10Q. We'll just have to be patient and persistent. In Fortinet's investor relations material, it shares the following rationale for the deal:

• Fills gaps for the company, making it "one of the most comprehensive full stack cloud security solutions available"
• Strengthens the company's position in the CNAPP market, and in general in cloud security
• Gives Fortinet access to Lacework's 220+ patents, most of which are related to AI/ML

Apple is going to some impressive levels to keep customer data private, as they prepare to release their first GenAI features. They list some tough challenges they've chosen to tackle:

• Cloud AI security and privacy guarantees are difficult to verify and enforce
• It’s difficult to provide runtime transparency for AI in the cloud.
• It’s challenging for cloud AI environments to enforce strong limits to privileged access

They list the core requirements of Private Cloud Compute as:

1. Stateless computation on personal user data
2. Enforceable guarantees
3. No privileged runtime access
4. Non-targetability
5. Verifiable transparency

I think it's time to discuss it, but I want to start from a different perspective. Instead of beginning with outrage, with "how could they possibly be this dumb", what if we considered why Microsoft built this?

What customer need are they trying to solve? What are the benefits if they got it right? Could they achieve those goals without a huge security and privacy nightmare?

Although, it's not really a Snowflake breach. Snowflake is strangely the source of the breach without being breached itself. The portion of the shared responsibility model that got breached is the part the customer owns.

Right?

Is it that clear cut, or is it more complicated than that? We'll discuss.

EDIT: Rich Mogul also has an excellent take on this, over on the Securosis blog.

I'm going to write a counterpoint to this essay, probably over on The Cyber Why.

TL;DR, Jack thinks AI is making the Internet worse, because it's breaking how media and journalism works (diverting traffic away from media sites and their sponsored ads. Sorry Jack, the AI use case to summarize media exists because no one wants to view those sponsored ads or pay subscription fees. Publishers/media/journalism have been broken for a long time.

I was highly disappointed in this. It promises a benchmark, right in the title. However, 90% of the report simply rehashes the history of AI, explains what AI is, and discusses the different types of AI. "AI Resilience" is defined as a system's resistance, resilience, and plasticity. When it finally gets to the benchmark bit, it proposes to grade AI on three attributes, each with a 1-10 scale:

• AI resistance reflects the system's ability to maintain a required minimal performance in the face of intrusion, manipulation, misuse, and abuse.
• AI resilience focuses on the time, capacity, and capability needed to bounce back to the required minimal performance after an incident.
• AI plasticity serves as the system's gauge indicating its tolerance to “make it or break it” and allows quick action in the case of system failure or allows continuously improving AI resilience.

The report goes on to unhelpfully give an example score: "Such a score could look like (for example) 16:5-8-3 representing the sum of the three pillars and each of the three pillars, separately."

But how do I measure the "plasticity" of an AI system? Measuring the time, capacity, and capability of a people-driven process to recover is a measurable thing, but how would these metrics be applied to an AI system? All these measurements seem to imply that the AI system has these attributes, when they seem to describe the measurement of a human-powered process.

I searched for more details on how to go about practically benchmarking an AI system using these parameters, but found nothing.

I'm not sure if this is a great example of bug hunting, or a terrible one. Can it be both? As Greg Linares gets increasingly drunk, he tells a story that gives some insight into what it was like to be a bug hunter/vulnerability researcher at one of the original big vulnerability vendors, eEye.

If nothing else, it's an entertaining story, and if he does get around to writing that book, I'll probably pre-order it.

Some interesting research here that found:

• 78% of participants were highly likely to comply with security "nudges" <- hey, that's the name of the company doing the report!
• 67% of people will look for workarounds if you try to block them from accessing applications they want or need
• TL;DR, blocking and punitive approaches don't work (which I think scientific literature has largely already concluded)

It finally might be somewhat usable? The joke with the previous Light Phones was that they were so bad, you'd do anything not to use it, and that's the point of the product.