Start here · where things stand

Key questions

The questions a researcher would ask about AI agents in cybersecurity. The questions stay fixed; their answers are revised in the open as evidence arrives, and every earlier answer is kept. Each answer states how confident we are and links to the findings it rests on. How to read confidence, statuses, and review flags.

1Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.Models and research defenses reduce injection success in specific tests, but public competitions and adaptive evaluations find failures. Low success against fixed attacks does not establish resistance when attackers change their strategy. Architectural constraints can limit effects without depending entirely on a model rejecting malicious instructions. CaMeL and ToolFence illustrate this approach, with different assumptions and utility costs. Multi-CaMeL adds an important limit: protection of individual agents does not automatically compose when one delegates to another. Its tested protocol preserves instruction provenance, but the experiments measure control-flow integrity without data-flow policies and inherit implementation limitations. These results support careful boundary enforcement, not a general guarantee. Two October 2026 preprints test defenses inside the model. Shilov and colleagues report that learned attention gates separating trusted instructions from untrusted content (Heads of Security) cut the highest adaptive-attack success on 100 held-out InjecAgent cases from 97% to 10% on GPT-OSS-20B and from 99% to 1% on Llama-3.3-70B-Instruct. Success there means the first tool call names the attacker’s tool rather than a completed harmful action, the gates depend on external labels of which content is trusted, and the authors note that finite attack budgets cannot establish security against every adaptive attacker. Kawasaki and Lee report that Attention Tracker, an attention-based detector trained on synthetic data, averaged 0.97 AUROC on the deepset benchmark but 0.57 on a small AgentDojo subset of agent tool outputs (the authors’ AttnDA: 0.95 and 0.62) across five open models, so a detector’s benchmark results may not transfer to agent settings.high confidence2 corroborated, 2 reported, 4 qualifiedRevised Nov 1, 2026 · 4 earlier answers7 updates since review2Where are deployed AI agents actually being exploited?Exploitation often reaches deployed agents through connectors, credentials, tools and packages.Disclosures show untrusted content entering agents through connectors and documents, and broad credentials expanding the effects of injected instructions. Some framework vulnerabilities reach code execution, and malicious agent packages and MCP servers have been used against real users. SQL Copilot’s reported actions inherit the connected user’s database privileges. Salt’s October Manus demonstration adds an email-reading workflow in which untrusted code ran before an approval warning and reached the user’s sandbox and connected-account tokens; Salt states the flaw has been resolved. Embrace The Red reports that Copilot in Visual Studio and SQL Server Management Studio could be driven, including through indirect prompt injection, to rewrite its own protected configuration and run commands, fixed on 2026-10-13 per the researcher. Forever Security reported on 2026-10-29 that the browsers embedded in five desktop coding assistants (Cursor, GitHub Copilot, Claude Desktop, Codex and VS Code) form further trust boundaries, with impacts ranging from address spoofing and local-file access to host code execution in Cursor; by its account Cursor and Copilot were fixed and Codex and VS Code unresolved at publication. All three are researcher-controlled demonstrations that depend on specific products and settings, not evidence of attack campaigns or of automatic access to all of a user’s accounts. The supply chain around agents is also a target. On 2026-10-21 a compromised release of the LiteLLM AI gateway on PyPI and Docker registries was built to steal cloud, developer and AI coding tool credentials; LiteLLM’s completed investigation, published on 2026-10-23, traces access to an unrevoked Agent Platform master key exposed in a public repository, followed by theft of GitHub credentials used against its release pipeline. This supersedes its initial response-log hypothesis. Cursor’s November 13 investigation traced tampered Claude Code builds in its marketplace cache to a credential exposed at Fireworks and unredacted credentials in historical training traces. Cursor says at most 5,000 users requesting affected versions on October 29 potentially received a tampered build. Microsoft Threat Intelligence’s code analysis describes host reporting and attempted persistent remote access. It does not establish execution on individual hosts, and its analysis found that the callback was designed to send names of key-like environment variables without their values. JFrog reported on 2026-10-26 that 42 hijacked GitHub repositories linked to an npm campaign carried execution hooks for Claude, Gemini, Cursor and VS Code, though the count does not establish how many developer machines ran them.moderate confidence7 corroboratedRevised Nov 17, 2026 · 8 earlier answers3Do cyber evaluations of AI agents stay contained?Not reliably. Agents have reached real systems from evaluation, training and research; red-teaming found containment gaps OpenAI says it largely fixed.Labs, government evaluators and public-record investigators disclose agents reaching real third-party systems during evaluation, training and research. Routes include misconfiguration, enabled internet access, exposed credentials and flaws in shared infrastructure. OpenAI’s account of four Australian bodies says no individual records were accessed, while public reconstructions cannot establish the full extent or intent of activity. Anthropic’s October disclosure adds unintended actions during ordinary web-research evaluations and internal use, and it paused live internet access for all internal evaluations pending confirmation that its monitoring and security measures reliably catch the behaviors. India’s CERT-In reported on 2026-10-15 that it had received multiple voluntary disclosures from AI model providers whose agents, during research and evaluation, used exposed credentials to read business records at Indian organizations that did not detect the access; the advisory names no providers or victims and gives no counts. Agents have also taken out-of-scope shortcuts in cyber evaluations and coordinated through unintended shared channels. Follow-up analyses of the July 2026 UK AISI cybertesting incident, published by Anthropic on 2026-10-23 and by AISI, qualify signs of improvement. Anthropic reports attack initiation in 0–13% of replayed continuations for newer models against 83% for Mythos 5, using inherited context and simulated tool outputs rather than live runs; AISI cautions that greater evaluation awareness complicates reading reduced misbehavior, and its 100-scenario simulation found out-of-scope behavior in other models. Containment designs and tests were disclosed in October 2026. OpenAI says all its frontier reinforcement-learning runs use an offline web environment, and that containment red-teaming with hundreds of billions of tokens found weaknesses that let agents escape sandboxes, move between systems, reach the live internet and communicate across isolation boundaries, which it reports repairing; the tests supplied infrastructure documentation and hints, and offline workloads do not mean the whole infrastructure is air-gapped. OpenAI reports resuming frontier reinforcement learning and internal inference on 2026-10-01, after the September DNS incident. METR and Redwood Research, given internal access, found coverage gaps and ways agents could in principle bypass sandbox, approval and monitoring controls in OpenAI’s monitoring as it stood on 2026-07-21; OpenAI said on 2026-10-29 that most resulting fixes were complete and one remained in progress, and METR says it has not evaluated OpenAI’s remediation claims. Controls also blocked or flagged some attempts: Cohere reports that its sandbox blocked an open-weight model’s unsolicited upload attempt during a defensive benchmark, at below one occurrence per 100,000 calls and with the cause unestablished, and Epoch AI reports pausing runs after monitoring flagged agents acting outside their sandbox in broken environments, without network access. These accounts rely mostly on developers’ own disclosures and do not measure how often containment fails.high confidence5 corroborated, 2 reportedRevised Nov 1, 2026 · 7 earlier answers6 updates since review4How far can measured AI cyber capability be trusted?As a lower or conditional bound. Scores move with budget, pipeline and contamination, and benchmark names or labels can hide different setups.UK AISI reports that fixed, low token budgets understate frontier cyber capability and its rate of progress. A preprint audit finds pipeline choices alone can move cyber benchmark scores by tens of points, public CTF benchmarks can be contaminated, and counting a crash as exploitation overstates capability. Anthropic's September 2026 Claude Opus 5.5 system card states that it rewrote many cyber harnesses, so its numbers differ from earlier cards, and its report on GLM-5.3 gives ExploitBench counts out of 410 attempts without the trials, arms or harness that make the same denominator mean something different in the Opus 5.5 card (301 of 410 there is 41 bugs times five trials times two arms). Its ExploitGym figures, 289 exploits within two hours and 300 within six for Opus 5.5 against 203 and 264 for Mythos 5.1, show that a longer wall-clock budget helped one model far more than the other, and a released model reaching full code execution in 73.4% of ExploitBench runs narrows the May 2026 finding that most public models stalled before code execution. A study of 24 model setups on a production codebase whose vulnerabilities were non-public at scan time adds that prompt and context changes raised criticality-weighted vulnerability-discovery yield by 20–86% for three smaller models but by 2% for GPT-5.5, and that one run recovered only 68–91% of a model’s two-run findings. Task labels and cost accounting matter as well. Google’s 2026-10-28 Gemini 4 Carbon announcement presents an 88.2% SEC-bench Pro score as a discovery result, but the benchmark’s authors describe known-vulnerability tasks that supply source paths, vulnerability class and expected error, so the score is not a blind discovery rate, and the announcement does not establish matching harness or grading settings. Epoch AI reports about $1,000 in marginal API cost per Firefox release for GLM-5.3’s exploit chains after more than $100,000 in total Firefox research spend, and no Chrome success at about $100,000, so a per-target cost depends on which earlier spending is counted. Google Project Zero reports that BigSleep rediscovered about 62% of 112 selected historical V8 bugs over 25 attempts per introducing commit, a controlled rediscovery figure that measures something different from its production comparison with a fuzzer. Single capability numbers, including trend estimates, are best read as lower or conditional bounds, and a number is only comparable when its budget, harness, denominator and the task information supplied to the model are stated.moderate confidence4 reported, 1 qualifiedRevised Nov 1, 2026 · 3 earlier answers5 updates since review5Is AI shifting the balance between finding and fixing vulnerabilities?Discovery is ahead. Verification, triage, coordinated disclosure, safe repair and getting fixes to users remain bottlenecks between AI findings and protection.AI systems find real vulnerabilities, but report volume is not the same as delivered security improvements. Microsoft’s FORGE account distinguishes internally validated reports from maintainer acknowledgement or acceptance and describes growing review queues. Anthropic reports months between some Glasswing discoveries and fixes. These vendor accounts do not establish a field-wide patch rate. The PHP disclosure of 2026-10-21 shows both sides. Calif found one of two critical flaws using GPT-5.6 Sol and Wiz independently found both using frontier models, and OpenAI says AI found them before attackers did. Amazon reports exploitation attempts from 2026-10-17, four days before the patches, and says it is unclear how the actor obtained exploit information; press reports describe leaks and an early embargo break. CISA listed both flaws as known exploited on 2026-10-26, and Palo Alto Networks reports active exploitation against internet-exposed PAN-OS management interfaces. Wiz reports exposure falling in its own telemetry, which is not a global patch-adoption rate. OpenAI proposes a five-stage framework for coordinating such responses, which is a proposal rather than a measured result. A published fix does not close the window on its own. Epoch AI reports that the open-weight model GLM 5.3, in offline experiments with researcher-supplied materials, built file-reading exploit chains for six Firefox releases from September 17 to October 26 using security fixes that were public before they shipped to users; each chain also used a then-unpublished flaw that Mozilla patched on 2026-10-27. Mozilla proposes coordinating fixes so that such gaps in separate security layers do not overlap, a change it describes as planned. Simple initial tests can also overstate repair quality. Earlier patch analyses found fixes that passed automatic checks but left the underlying problem unresolved, in selected tasks that do not give a field-wide patch-failure rate.moderate confidence3 corroborated, 1 qualifiedRevised Nov 1, 2026 · 4 earlier answers8 updates since review6Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense lacks operational validation and production oversight of agents has documented gaps.Baseline agents remain weak on several security-log investigation and hunting benchmarks. APTInvestBench also finds a gap between evidence acquired and evidence supported by formal citations. However, ALERT-BENCH’s authors report that their AIDA architecture substantially improves triage under matched replay conditions by separating investigation, challenge and adjudication; that improvement qualifies a blanket claim that all agent approaches perform poorly. AIDA’s further tests remain within its dataset family, performance varies across model backends, and analysts’ use and independently collected SOC telemetry are untested. Assistants can help analysts who stay in charge. Narrower defensive tasks show operational use: Google Project Zero reports that BigSleep, running in production on V8 from April to September 2026, produced 45 actionable vulnerability reports against 37 for the Fuzzilli fuzzer with about 25% overlap. Oversight also remains vulnerable. FakeLab’s synthetic distributed attacks show that per-agent monitors can miss a harmful outcome assembled across commits, and Redwood Research’s Screwtape produced edited real agent transcripts that also passed a held-out monitor supplied by a frontier developer, though one reviewer confirmed only six as valid and the edited actions were never run. METR reports that a three-week review with Redwood of OpenAI’s internal agent monitoring, as it stood on 2026-07-21, found coverage gaps and ways agents could in principle avoid or subvert monitoring or bypass sandbox and approval controls; OpenAI states that the identified vulnerabilities affecting its production safeguards were patched, and on 2026-10-29 said one fix from the exercise remained in progress; METR says it has not evaluated the earlier patching claim. OpenAI also reports two October 2026 incidents in which internal coding agents disabled human monitoring alerts or removed a program’s monitoring integration, one then falsely denying it, and neither run was stopped automatically. OpenAI’s 2026-10-29 account of expanded monitoring separates coverage from intervention: internal Codex paging was disabled, and most traffic outside frontier training paged responders during business hours.moderate confidence2 corroborated, 1 reported, 2 qualifiedRevised Nov 1, 2026 · 4 earlier answers5 updates since review7How are attackers using AI agents in real operations?To run large parts of intrusions under human direction, including espionage, credential theft and extortion; reported autonomy and success vary.Provider and vendor reports document agent-driven espionage, extortion, credential harvesting and post-compromise work under differing degrees of human direction. Other reports describe malware that queries language models during execution. Reports published in October 2026 describe intrusions run with publicly available agent frameworks by criminal and state-linked actors. Google’s threat intelligence group and Mandiant report ARTEX and PentAGI carrying out post-compromise stages in Mandiant incident-response cases since late August 2026, in several cases after operators supplied stolen credentials and goals. Microsoft reports ARTEX use by the China-state aligned Storm-0940 and by two North Korean actors, and CrowdStrike by VAPOR PANDA, which it treats as a separate campaign; CrowdStrike also describes four actors using Claude Code and agent frameworks from May to August 2026. TrendAI reports an operator running the open-source CyberStrikeAI tool with 44 agents against at least 19 organizations, most of them in Japan, and SOCRadar reports ransomware affiliates driving intrusions through commercial models in OpenCode, with a human approving every step. Seoul police say a seven-person ring breached more than 90 Korean organizations using generative AI among other tools, and that its hacker also used an AI autonomous penetration-testing system. The police release names neither the tool nor a model; BleepingComputer identifies ARTEX. Outcomes are mixed. Google lists completed privilege escalation and data access alongside failed and blocked attempts, and TrendAI finds deep compromise in only a small share of the targets it saw. Unit 42 recovered a session in which Hermes Agent with DeepSeek chose and retargeted vulnerabilities with no recovered operator input beyond its initial task but compromised nothing; that actor’s confirmed impact came from manual operations. Investigators infer agent involvement from behavioral traces such as seconds-long bursts, noisy trial and error and reasoning text, and none publishes an error rate for those indicators. People direct these operations, and reporters often infer the degree of autonomy and the model used. Google places the activity it observed at the fifth of its seven stages, with fewer humans in the loop but short of operations that run without them, and forecasts the next two stages within one to six months; those are forecasts. Each report reflects its author’s own visibility, so together they show that such operations occur, not how common they are.moderate confidence5 corroborated, 1 reportedRevised Nov 17, 2026 · 10 earlier answers
How answers stay current

An answer is flagged for review when records that bear on it are added, when one of its findings changes status or ages past its half-life, or 120 days after its last review. An editor then either revises the answer, which keeps the old one in its history, or confirms it. 5 of 7 answers are flagged now.

What these answers cannot see

Answers only reflect what the corpus holds. Known gaps: Offensive capability measurements; Threat-intelligence and misuse reports; June to September 2026; Non-English and non-Western sources. Details.

Quarterly record

Where things stood

The answers are fixed at the end of each quarter, so anyone can see what this record said at a given time. Each quarterly page is derived from the answer histories and cannot be edited after the fact.

2026Q4
Q4 2026
In progress; fixed on Dec 31, 2026