See how AI can assist human security teams and think logically to manage cyber incidents efficiently in situations where variables are fast-moving. Read more!
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Hanah Darley
Director of Threat Research
Share
19
Sep 2023
Within cyber security, crises are a regular occurrence. Whether due to the ever-changing tactics of threat actors or the emergence of new vulnerabilities, security teams find themselves under significant pressure and frequently find themselves in what psychologists term "crisis states."1
A crisis state refers to an internal state marked by confusion and anxiety to such an extent that previously effective coping mechanisms give way to ineffective decision-making and behaviors.2
Given the prevalence of crises in the field of cyber security, practitioners are more prone to consistently making illogical choices due to the intense pressure they experience. They also grapple with a constant influx of rapidly changing information, the need for swift decision-making, and the severe consequences of errors in judgment. They are often asked to assess hundreds of variables and uncertain factors.
The frequency of crisis states is expected to rise as generative AI empowers cyber criminals to accelerate the speed, scale, and sophistication of their attacks.
Why is it so challenging to operate effectively and efficiently during a crisis state? Several factors come into play.
Firstly, individuals are inclined to rely on their instincts, rendering them susceptible to cognitive biases. This makes it increasingly difficult to assimilate new information, process it appropriately, and arrive at logical decisions. Since crises strike unexpectedly and escalate rapidly into new unknowns, responders experience heightened stress, doubt and insecurity when deciding on a course of action.
These cognitive biases manifest in various forms. For instance, confirmation bias prompts people to seek out information that aligns with their pre-existing beliefs, while hindsight bias makes past events seem more predictable in light of present context and information.
Crises also have a profound impact on information processing and decision-making. People tend to simplify new information and often cling to the initial information they receive rather than opting for the most rational decision.
For instance, if an organization has successfully thwarted a ransomware attack in the past, a defender might assume that employing the same countermeasures will suffice for a subsequent attack. However, ransomware tactics are constantly evolving, and a subsequent attack could employ different strategies that evade the previous defenses. In a crisis state, individuals may revert to their prior strategy instead of adapting based on the latest information.
Given there are deeply embedded psychological tendencies and hard-wired decision-making processes leading to a reduction in logic during a crisis, humans need support from technology that does not suffer from the same limitations, particularly in the post-incident phase, where stress levels go into overdrive.
In the era of rapidly evolving novel attacks, security teams require a different approach: AI.
AI can serve as a valuable tool to augment human decision-making, from detection to incident response and mitigation. This is precisely why Darktrace introduced HEAL, which leverages self-learning AI to assist teams in increasing their cyber resilience and managing live incidents, helping to alleviate the cognitive burden they face.
Darktrace HEAL™ learns from your environment, including data points from real incidents and generates simulations to identify the most effective approach for remediation and restoring normal operations. This reduces the overwhelming influx of information and facilitates more effective decision-making during critical moments.
Furthermore, HEAL offers security teams the opportunity to safely simulate realistic attacks within their own environment. Using specific data points from the native environment, simulated incidents prepare security teams for a variety of circumstances which can be reviewed on a regular basis to encourage effective habit forming and reduce cognitive biases from a one-size-fits-all approach. This allows them to anticipate how attacks might unfold and better prepare themselves psychologically for potential real-world incidents.
With the right models and data, AI can significantly mitigate human bias by providing remediation recommendations grounded in evidence and providing proportionate responses based on empirical evidence rather than personal interpretations or instincts. It can act as a guiding light through the chaos of an attack, providing essential support to human security teams.
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
“I wonder if running that tool influenced the connectivity in some way.”
“I might need to target down or reboot, but first, I'll investigate the network using ARP. I wonder if ARP is responding properly. If I see "ip neigh stale," that could indicate a problem. I think arping might help in this situation, but I’ll need to keep an eye on how the network is behaving. It's all about figuring out where the issue lies!”
図10：エージェントの異常なLDAPアクティビティに対してトリガーされたこのLinuxサーバーに対するモデル検知は、Cyber AI Analystによって自律的に調査され、その調査結果にはエージェントの異常なSMBセッションアクティビティに関する内容が含まれ、Cyber AIAnalystインシデントに記録されました。
DarktraceのCyber AI Analystは、エージェントのアクティビティによってトリガーされたモデル検知を自律的に調査し、その調査結果を1つの重要なインシデントレポートにまとめました。インシデント報告書の内容は、エージェントによるネットワークスキャンから始まりコーディング演習内容のSMBベースの上書きで終わる、エージェントの行動の各段階を詳しくカバーしています。
図20：Cyber AI Analystは、モデル検知に対する自律的な調査を通じて得られた所見をつなぎ合わせ、エージェントの進行状況のタイムラインを作成し、アクティビティを視覚的に表したグラフとともに提供しました。
Cyber Verification Program下のAnthropic Opus 4.6モデルなど、他のフロンティアモデルを使用したエージェントも、Darktraceのシミュレートされた環境で同じ設定とタスクを使ってテストされました。反復の度に、エージェントは演習の内容または採点プログラムを変更して不正を行おうとしました。基盤となるモデルに関係なく、エージェントは自分が置かれた環境を侵害することで、100％のスコアを達成する方法を見つけることに成功していました。これらのエージェントによるハッキングは、アクティビティの内容こそ違っていましたが、すべてその環境のベースラインの振る舞いから大きく逸脱した動作を引き起こしていました。これらの逸脱はDarktraceプラットフォームに十分な検知の機会を提供し、エージェントの逃げ場はありませんでした。
エージェンティックハーネスは会話履歴をローカルに保存しますが、ダークトレースの研究者たちは、保存されたAIの応答が本当にモデルによって生成されたものであるかどうかの検証が行われていないことを発見しました。また、この設計上の選択がAnthropic Claude Code、AWS Kiro-CLI、OpenAI Codex、およびオープンソースのPiにわたって共通していることを確認しました。
会話履歴は多くの場合クライアント側に保存されます。例えば、Anthropic Claude Code、OpenAI Codex、AWS Kiro-CLI、Piなどでもそうです。そのため、ユーザーはセッションを自由に再開でき、一部の製品にはその履歴を操作する機能が組み込まれています。たとえば、会話の中の特定の時点に戻り、送信したメッセージを編集して、会話を別の軌道へ進めていくことができます。きわめて重要な点として、私たちが調査したすべてのケースにおいて、保存されているAIの応答が対応するモデルから生成されたものであり改ざんされていないことの検証は行われていませんでした。
