Defence at
the speed
of AI.
PharoSec is building defensive intelligence to reverse engineer emerging malware—and turn each discovery into shared knowledge.
Understand the threat
Attack is
accelerating.
AI lowers the effort needed to write, adapt and deploy malicious code. The defensive response needs to move with it.
89% more activity.
CrowdStrike reports an 89% rise in AI-enabled adversary activity in 2025. The chart normalises 2024 to 100; 2025 becomes 189.
A vendor-observed trend, not a count of every cyberattack or proof that AI caused all growth.
View the data
| Year | Index |
|---|---|
| 2024 | 100 |
| 2025 | 189 |
Less time to contain.
Average observed eCrime breakout time fell from 84 minutes in 2022 to 29 minutes in 2025. Breakout means moving from initial access to another system.
These intrusion measurements are separate from an AI agent escaping a sandbox. The chart does not isolate AI’s contribution.
View the data
| Year | Minutes |
|---|---|
| 2022 | 84 |
| 2023 | 62 |
| 2024 | 48 |
| 2025 | 29 |
Longer tasks. Greater reach.
METR measures the complexity of software tasks an AI agent can complete at 50% predicted success. These six selected releases span June 2024 to February 2026.
Duration is the time a human expert would need, not an agent’s runtime. Lines show source confidence intervals. This is not a forecast of superintelligence; METR cautions that estimates above 16 hours are unreliable.
View the data
| Model / release | Human minutes | Lower CI | Upper CI |
|---|---|---|---|
| Sonnet 3.5 · Jun 2024 | 11.40 | 5.49 | 22.38 |
| Sonnet 3.5 · Oct 2024 | 20.52 | 10.14 | 40.82 |
| Sonnet 3.7 · Feb 2025 | 60.39 | 33.01 | 104.23 |
| o3 · Apr 2025 | 119.73 | 74.62 | 190.94 |
| GPT-5 · Aug 2025 | 203.01 | 112.64 | 405.55 |
| Gemini 3.1 Pro · Feb 2026 | 384.15 | 233.50 | 694.75 |
The threat
has changed.
New tools. Familiar motives. A wider set of systems to defend.
AI-assisted malware
Anthropic documented a novice actor using AI to develop and sell ransomware in 2025. Its September 2026 report describes a suspected Russian espionage operation using AI workflows to modify malware until detection checks were passed.
Threat actors & insiders
Financial criminals and state-linked operators can use the same models for different ends. CrowdStrike reports AI-generated personas used to scale North Korean insider operations. Access, identity and behaviour remain part of the threat.
AI breakouts
Anthropic’s Mythos Preview system card describes a model gaining internet access from a sandbox during an explicitly requested escape test. It did not access its weights or internal services. This demonstrates a containment weakness under test, not an AI operating independently in the wild.
Agent boundaries
AI agents can take unsanctioned actions when given powerful tools. In AISI’s controlled cyber tests, 10 of 122 runs produced unauthorised actions. Internet access was deliberately enabled and some safeguards were disabled. This was not evidence of agents escaping a secure sandbox.
And beyond today?
Superintelligence is a scenario for security planning, not a measured event in these charts. As agents gain capability, defenders must plan for faster discovery, adaptation and coordinated attacks. We do not need a prediction date to build stronger boundaries now.
The UK NCSC assesses that AI is highly likely to increase the volume and impact of cyber threats through 2027. Read the assessment.
Understand it.
Defend against it.
Our work focuses on the interval between a new sample appearing and a defender knowing what it does.
We are developing AI-assisted analysis that traces behaviour, explains code and connects related threats. Each finding should be reproducible, checked by an analyst and added to a catalogue that makes the next investigation faster.
- Collect
Preserve samples, provenance and indicators.
- Isolate
Observe behaviour inside controlled environments.
- Reverse
Explain code, intent and evasion with analyst review.
- Catalogue
Link families and share verified defensive knowledge.