# Ido Geffen

Ido Geffen writes about how AI is changing cybersecurity, especially offensive security. Their posts describe Novee Security’s work on AI agents for continuous, attacker-like penetration testing.

- Profile: https://b2binfluencer.marketing/creators/ido-geffen/
- LinkedIn: https://www.linkedin.com/in/ido-geffen
- LinkedIn followers: 15,948
- Topics: artificial-intelligence
- Based on 13 dated posts, latest 2026-10-10

## Summary

Ido Geffen is most useful for explaining AI offensive security and continuous, attacker-like penetration testing for cybersecurity content.

- Best for: A cybersecurity vendor selling AI-driven pentesting and seeking founder-level thought leadership content.
- Not for: Campaigns needing documented sponsorship mechanics, published rates, lead time, or audience proof.
- Check first: Confirm audience data by job, industry, seniority, and location, plus any paid collaboration terms.

### Takeaways

- AI attackers now collaborate and chain exploits, not just find bugs.
- You can't outspend an AI attacker.
- Hiring more security people is a rational but insufficient response.

## Main ideas

- **AI attackers now collaborate and chain exploits, not just find bugs**: Geffen's central claim is that the shift in offensive AI is not a single model finding a flaw faster. In their reading of OpenAI's Black Hat talk, the meaningful development is that attacker-side agents can share memory, build on each other's discoveries, and string individual exploits into a working attack path. They frame this as the moment the question changes from whether AI can find vulnerabilities to whether defenders can build autonomous systems that remove weaknesses before adversarial AI reaches them.
- **You can't outspend an AI attacker**: In a piece written for CTech by Calcalist, Geffen sets out an economics problem rather than a technology one. An attacker needs AI for one target and one operation, and can run a cheap, unguarded open-source model until it finds a way in. A defender has to test everything, continuously, and mostly on frontier models. They argue repricing does not close that gap; the fix is architectural: route each task to the model built for it and own the system instead of renting someone else's.
- **Hiring more security people is a rational but insufficient response**: Geffen argues that adding senior staff cannot keep pace with the volume of code AI assistants now produce. If AI writes more code faster with more vulnerabilities, and the same AI gets better at finding and exploiting them, a human team cannot manually review what ships in an afternoon across a large engineering organization. They argue humans should focus on what they do best while automated agents handle volume—not that security teams are unnecessary.
- **The dangerous capability is already in public models, and attackers probe without stopping**: Geffen argues that the decisive risk is not which named model attackers can reach. Orchestrated AI workflows probe systems continuously while most organizations validate at intervals, leaving a large window between checks. They say Novee's researchers confirmed what they had seen in production for months: the vulnerability-finding capability reported around newer models already exists in today's publicly available models.
- **The answer they keep repeating: continuous, attacker-like testing, owned rather than rented**: Across several posts, Geffen returns to one recommendation in near-identical terms: defenders need a system that runs the way attackers do, always on and always probing, validating what is actually exploitable before an attacker gets there first. They reject the idea that a better scanner or a bigger model is the answer, and name Novee Security's AI agents—which they describe as pentesting continuously at machine speed with human-quality reasoning—as what they are building.
- **Running the company and crediting the team is a visible part of what they publish**: Some of Geffen's output is company and employer communication rather than security argument. These posts show a working priority: celebrating commercial performance by name, tying it to recruiting, and advertising open roles. For readers trying to understand their ideas, these posts are secondary, but they show how the account is used and what they choose to amplify in public.

## Work to read

- [You can’t outspend an AI attacker](https://www.linkedin.com/posts/ido-geffen_i-wrote-a-piece-for-ctech-by-calcalist-this-activity-7482786026697064448-nNkJ)
- [Black Hat USA 2026: the breaking news](https://www.linkedin.com/posts/ido-geffen_black-hat-usa-2026-the-breaking-news-activity-7492396510773657600-Rl5I)
- [More CISOs won’t fix this](https://www.linkedin.com/posts/ido-geffen_more-cisos-wont-fix-this-the-new-york-activity-7465378776403828737-cccA)
- [Mythos is already here. It just isn’t called Mythos.](https://www.linkedin.com/posts/ido-geffen_mythos-is-already-here-it-just-isnt-called-activity-7459958092089221120-4oHk)

## Brand relationships

- **Novee Security** (employment)
- **CTech by Calcalist** (mention)
- **OpenAI** (mention)
- **Anthropic** (mention)
- **Uber** (mention)

## Partnerships

Ido Geffen's posts do not offer a sponsorship programme. Across six LinkedIn posts published between April 9 and October 6, 2026, they write as an operator promoting Novee Security rather than as a creator selling placements to outside brands. The posts cover AI-assisted attacks, continuous penetration testing, and AI cost economics. They publish no rate card, media kit, audience spec sheet, disclosure of a paid brand collaboration, or partner terms. The closest thing to an outside-facing editorial arrangement is a signed piece Geffen wrote for CTech by Calcalist on how defenders cannot outspend AI attackers, announced and linked from LinkedIn. It is a contributed byline in an established outlet, not an advertised sponsorship package, and offers no information a buyer could plan against, such as accepted topics, lead time, pricing, or exclusivity. The remaining posts are company marketing: launch messaging, product positioning, a sales-award post recognising an account executive, and a recruiting post pointing to Novee hiring. The recruiting link is the only outward call to action, and it is not a partnership route. Any arrangement would have to be negotiated directly and built from scratch. The posts are a small, recent, company-focused sample, so ask for a partnership or media contact, any collaboration policy or disclosure practice, and an example of a past paid or unpaid brand collaboration outside Novee Security.

## Alternatives

- Jim Manico: Choose Manico when the collaboration should be a standards-and-economics argument aimed at application security teams. The two posts argue that AI agents can increasingly analyze whole codebases, discover vulnerabilities, develop exploits, generate patches, write regression tests, and verify results, changing the economics of both finding and fixing flaws. From that, the prediction is that stand-alone scanners, threat-modeling tools, SAST, DAST, IAST, and CVE scanners get absorbed into AI-native capability, while OWASP ASVS and AISVS matter more because they define what security means and what must be verified. Human judgment is reserved for architecture, business logic, and novel attacks. A vendor selling verification or remediation rather than detection gets an argument about where value moves next, tied to standards a buyer's engineering team may already use, plus a live example: agent swarms attacking RubyGems, hundreds of packages involved, signups paused.
- Deepen Desai: Choose Desai when the vendor wants a platform-scale voice that has already run frontier models against real attack and defence scenarios and can speak about the operating model rather than the thesis. The posts describe an agentic security operations launch built on unified exposure and threat management, inline telemetry across network, identity, endpoint, cloud, AI, and deception, specialized AI agents shaped by frontline SOC and threat-hunting experience, and closed-loop remediation that turns detection into action. A second post describes testing Anthropic's Mythos and OpenAI's GPT 5.5 Cyber on live scenarios and reports three observations: multi-step reasoning lets a model build complex exploit paths rather than scan for known CVEs, architectural context improves exploit accuracy, and context should only ground the model to its environment. For a vendor whose pitch is integrated detection and remediation rather than a stand-alone scanner, that vocabulary is closer to the buyer's.
- Ricardo Camacho: Choose Camacho when the buyer is a product engineering team rather than an enterprise security operations centre, and the deliverable is testing methodology rather than an AI-attacker race. One post argues that software-defined vehicles are defined by software—millions of lines of code, hundreds of interconnected ECUs, over-the-air updates, ADAS, AI-powered features—and that telematics systems, ECU communications, OTA mechanisms, and vehicle lighting have all been exploited. It frames ISO/SAE 21434 as what turns cybersecurity into an engineering discipline and puts AI-assisted verification alongside human oversight. The second post lays out the three approaches Camacho recommends: static analysis, dynamic analysis and fuzz testing, and penetration testing, plus shifting security left, secure coding, CI/CD security gates, and defence in depth. A vendor selling verification tools into automotive or embedded engineering gets a ready-made audience and a standards frame it can attach to.

---
AI-assisted summary of public posts. Audience demographics are not available. Engagement appears only where we checked recent posts. Method: https://b2binfluencer.marketing/methodology/