AI agents in cybersecurity, on the record
A sourced chronicle of how AI agents are used to attack, how they are measured and gated, how they are defended, and how they fail. Every claim links to its evidence, carries a status that changes as new work arrives, and feeds a list of open research questions.
- Records
- 342
- Sources
- 517
- Findings
- 57
- Key questions
- 7
- Methods
- 31
342 records from Feb 1, 2023 to Nov 17, 2026: 116 attack, 25 capability, 137 defense, 64 policy. The chronicle lists every record.
Key questions
The lasting questions about the field. Each answer is revised in the open as evidence arrives, and shows how confident we are and how solid its findings are. What the statuses and flags mean.
Week of Nov 16–22, 2026
Recent landmarks
Chronicle →Findings that moved
Ledger →Status changes caused by new evidence. Our own 14 corrections are in the changelog.
Multi-CaMeL narrows the single-agent guarantee: trust and provenance must be preserved across delegation, and the tested successor provides control-flow integrity under explicit assumptions rather than general data-flow protection.
ALERT-BENCH identifies large baseline triage failures but reports substantial improvement from AIDA’s separate challenge and adjudication. The weak-performance claim is narrowed to baseline approaches; operational reliability remains unestablished.
Public red-teaming competitions on 2025 and 2026 frontier models with built-in safeguards report much lower per-model success (0.5% to 8.5% in 2026), though every model was hijacked at least once. The substantial rates describe 2024 models and benchmarks.
ThreatDown independently documents a botnet whose implant is an agent framework driven by a model.
OpenAI reports self-generated instructions in compaction summaries.
Research openings
How much of measured cyber progress is measurement?
How much do cyber capability trends change when token budget, pipeline, and cheating controls are held constant?
FID-075What does a human approval actually check?
When an agent asks for approval, do people have the information and time to catch the consequential action?
FID-076What evidence should gate an AI-generated patch?
Which independent checks change the decision to accept an AI-generated patch, and what do they cost?
FID-088Reading paths
Prompt injection in tool-using agents
From the first description of indirect prompt injection to why published defenses fail against adaptive attackers, and what design choices hold up.
Evaluation validity · 5 stepsCan we trust cyber evaluations?
Capability claims drive release decisions. These records show how budgets, pipelines, cheating, and leaking environments change what evaluations report.
Vulnerability repair · 5 stepsFinding and fixing vulnerabilities with AI
AI systems now find real vulnerabilities at scale. Whether their fixes work is less clear.
Changelog
- Nov 17, 20262 key questions added or revised
- Nov 17, 20266 records revised
- Nov 17, 2026Finding reconfirmed: Separating an agent’s control flow from untrusted data can give provable protection against specified forms of hijacking under explicit assumptions.
- Nov 17, 20262 corrections to findings
Records summarize their sources in our words and link to them. Marks of significance () are editorial judgments. Findings carry a status that we change when new evidence arrives, and every record shows whether a Fide editor has reviewed it. What every label means · Methodology.