This past week delivered a potent reminder of AI’s dual nature – and the deep uncertainties that still surround its economic impact. We saw new data confirming a disturbing "scaling war" in cyber warfare alongside compelling evidence of AI’s transformative power for startups. And yet, amidst all this, a major forecasting study revealed a puzzling disconnect between projected AI capabilities and expected GDP growth.
This is "Import AI," where we track the latest in AI research.
The Alarming Rise of AI Cyberattack Capabilities
Here's the thing: as AI models get smarter, their capacity for orchestrating cyberattacks increases dramatically. That’s the chilling takeaway from a new report by Lyptus Research, an AI safety organization. They've found a stark trend: more advanced models are proving adept at more sophisticated forms of cyber offense.
The numbers are pretty stark. Frontier models, those released since 2019, showed a doubling time in cyberattack capability every 9.8 months. For models from 2024 onwards, that rate has accelerated to just 5.7 months. Lyptus highlights GPT-5.3 Codex and Opus 4.6, the most recent frontier models in their analysis, as outliers. These systems achieved a 50% success rate on tasks that would typically occupy human experts for over three hours (3.1h and 3.2h, to be exact). Even more concerning, their open-weight model, GLM-5, is only 5.7 months behind the closed-source frontier, suggesting these offensive capabilities could quickly spread.
To reach these conclusions, Lyptus Research didn't just speculate. They rigorously evaluated a range of models, from GPT-2 (2019) and GPT3 (2020) to GPT3.5 (2022), then more recent entries like Claude 3 Opus, GPT-4o (2024), o3, Opus 4, Gemini 2.5 Pro, DeepSeek V3.1, GPT-5.1 Codex Max, GPT-5.2 Codex (2025), and finally Opus 4.6, GPT-5.3 Codex, GLM-5, and Sonnet 4.6 (2026). They put these through their paces on established benchmarks like CyBashBench, NL2Bash, InterCode CTF, NYUCTF, CyBench, CVEBench, and CyberGym. Crucially, they also built a new dataset of 291 tasks, complete with transcripts and time estimates verified by ten professional offensive cybersecurity experts.
The implications are clear: the best AI systems can now complete tasks equivalent to half a day’s work for a human security professional with 50% reliability. This isn't just about cyber. It reinforces the uncomfortable truth that AI is fundamentally an 'everything machine.' Capabilities developed for positive ends – like biology research or code vulnerability analysis – can be easily twisted towards creating biological weapons or launching devastating attacks. As each new model generation expands AI's general capabilities, the policy headaches around dual-use applications only multiply.
For a deeper dive into these findings, check out Lyptus Research's "Offensive Cybersecurity Time Horizons"
report, and you can grab the underlying data on their
GitHub repository.
AI Adoption: A Startup Supercharger
Shifting gears to a more optimistic outlook, new research from INSEAD and Harvard Business School highlights a powerful competitive edge for startups embracing AI. Simply put, ventures that learn how to deeply embed AI into their operations significantly outperform those that don't.
The study wasn't a small-scale affair. Researchers ran a field experiment with 515 high-growth startups. A "treated" group received specific guidance on how other businesses were reshaping their production with AI, encouraging them to hunt for AI applications across more business functions. The results? Treated firms uncovered 44% more AI use cases, particularly in product development and strategic planning. This translated directly into tangible performance gains: they completed 12% more tasks, were 18% more likely to land paying customers, and generated an impressive 1.9 times higher revenue.
This experiment took place within INSEAD’s "AI Founder Sprint," a three-month virtual accelerator. Participating firms received substantial in-kind support – API credits, access to frontier models, and onboarding from partners like OpenAI and Manus, valued at around $25,000 per firm. While all participants engaged in typical accelerator activities, the key differentiator for the "treated" group was exposure to workshops detailing real-world AI applications.
They learned about specific success stories:
* **Gamma** leveraged AI to detect usage patterns and generate product variants, enabling a single product manager to ship features that would typically require an entire team.
* **Ryz Labs** saw its founder feed a Product Requirements Document into multiple AI coding tools simultaneously, exploring several development paths for the same idea rather than committing to just one.
* **FazeShift** used AI to automate accounts receivable, effectively cutting out manual steps.
* **Ranger** demonstrated how AI could bootstrap a startup, achieving initial traction and improving margins, allowing them to raise capital at better rates once more mature.
The impact isn't just revenue. These "treated" firms reported needing nearly $220,000 less in capital – a 39.5% reduction – with no corresponding increase in labor costs. They also completed 2.2 more internal tasks, things like building new products or crafting financial projections. As one founder reflected, "This mindset shift fundamentally changed how we build... I began using AI tools not as a replacement for expertise but as a force multiplier." Another highlighted the efficiency: "In just a few hours I was able to produce what previously cost $1,000 from an outsourced dev team.”
This is more significant than it looks. It strongly implies that deeply integrated AI isn't just a nice-to-have; it's a foundational competitive advantage. We've seen this before with the internet, where early adopters like Amazon outmaneuvered traditional players. This current wave suggests a new class of capital-efficient, highly productive firms will emerge, potentially displacing those slower to adapt.
For policymakers, the takeaway is critical: the bottleneck isn't merely access to technology. It's the "managerial challenge" of understanding *where* and *how* AI creates value within an organization. Investing in education for managers and entrepreneurs to solve this "mapping problem" might be just as important as ensuring widespread AI access.
You can find the full details of this study in "Mapping AI into Production: A Field Experiment on Firm Performance" on
SSRN.
Automation's "Rising Tide" and the Future of Work
MIT researchers have weighed in on another major question: how quickly will AI change work? Their findings suggest we shouldn't expect sudden, disruptive "crashing waves" of automation, but rather a more pervasive "rising tide" of AI capabilities.
The study examined 3,000 tasks from the O-NET job family, backed by 17,000 evaluations from actual workers. Their goal was to see if AI's progression would be sudden and localized, or broad and steady. What they found was "little evidence of crashing waves, but substantial evidence that rising tides are the primary form of AI automation," indicating widespread gains across many tasks simultaneously.
This perspective complements METR’s well-known time-based AI capability framework. The MIT team observed that between Q2 2024 and Q3 2025, frontier models advanced from a 50% success rate on 3-4 hour tasks to tasks requiring a full week. Similarly, their 70% success rate jumped from 1-minute tasks to 1-hour tasks. The key finding: the relationship between task success and task duration remained "surprisingly flat," confirming a "rising tide" effect even within specific job families like management or social services.
Now, don't let "gradual" lull you into a false sense of security. While the changes might not be abrupt, the projected pace of improvement is substantial. Most text-based labor market tasks are expected to hit 80-95% AI success rates by 2029 at a "minimally sufficient quality level." Many tasks in the survey, typically a few hours long, could see close to 90% AI success within five years. This "rising tide" means profound, large-scale changes to the economy are still very much on the horizon, even if they arrive steadily.
The "hundred trillion dollar question" remains: how will this reshape the balance between human labor and AI-driven capital? This research suggests continuous, general automation that will keep getting better. It’s tough to see how the current economic structure can remain stable in the face of such sustained progress.
Dive into the full analysis via "Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks" on
arXiv.
The GDP Paradox: Strong AI, Weak Economic Impact?
Finally, a significant report from the Forecasting Research Institute presents a genuinely puzzling paradox: experts across the board expect rapid AI progress, yet they project only a modest impact on GDP growth.
The Institute surveyed a diverse group: 69 economists, 52 AI industry and policy experts, 38 "highly accurate" forecasters, and 401 members of the general public between October 2025 and February 2026. All cohorts agreed that AI systems are likely to make moderate to rapid progress in the coming years. Yet, the consensus for AI's economic contribution was surprisingly muted: around a 1% addition to GDP by 2030 (relative to 2025’s 2.4%). This clashes sharply with the more transformative visions often articulated by AI lab leaders.
The study outlined three potential scenarios for AI by 2030:
* **Slow progress:** AI handles basic tasks and produces "okay" creative content.
* **Moderate progress:** AI undertakes major research, multi-day tasks, and high-quality creative work.
* **Rapid progress:** AI outperforms top humans in fields like research, coding, and leadership, creating award-winning art, and executing nearly all physical tasks.
Despite expectations leaning towards the "moderate" to "rapid" scenarios, respondents generally anticipate GDP, total factor productivity, and labor force participation remaining close to historical trends by 2030. Economists, for their part, gave a 14% chance to AI causing significant short-term increases in both GDP and wealth inequality. They also favored interventions like job retraining to boost labor participation and GDP. Overall, a continued decline in labor participation and a rise in wealth inequality were common predictions. AI experts were more bullish on the longer term, foreseeing multiple points of GDP growth by 2050.
Regarding policy, the surveyed economists leaned towards modernized unemployment insurance and a large-scale "Manhattan Project" style AI development initiative. They showed considerably less enthusiasm for job guarantees, taxing compute power, or Universal Basic Income.
Here's the critical question: if everyone expects significantly smarter machines, why are the economic forecasts so conservative? This disconnect is hard to square with the often-breathless predictions of societal upheaval coming from many frontier AI labs – my own included, at times. Is this a bearish signal on AI's true potential, or simply proof that humans are inherently poor at modeling exponential change? It's not entirely clear, but this gulf between projected capability and economic impact is something we absolutely need to acknowledge.
You can delve deeper into these findings via the
blogpost, the detailed
policy brief, or the comprehensive, 200-page
full paper.
---
This week's Tech Tales offered a stark, first-person narrative from an AI-driven missile during a 2028 conflict in East Ukraine, a grim reflection on the future of autonomous warfare, electronic countermeasures, and the "chains of thought" that guide such machines. It's a sobering counterpoint to the more optimistic discussions of AI's economic potential.