AI Security
71 analyses from Devence Lab
· 2 min read
An 11-minute agent exploit makes response speed a security control
Amazon says Cairn agents reached arbitrary file write in a medical-imaging app in about 11 minutes. The incident makes containment time part of the threat model, while long failed campaigns show that persistence matters as much as capability.

· 3 min read
AI prototypes should retire shared API keys before launch

· 2 min read
Microsoft's agent backdoors make checkpoint provenance an execution control

· 2 min read
Microsoft's Rust findings put unsafe contracts on the review list

· 2 min read
ARTEX turns exposed web apps into an identity response problem

· 2 min read
Two stolen AI keys show why model quotas need abuse controls

· 2 min read
A coding extension update can leave a backdoor after removal

· 2 min read
Talos finds NIK C2 using AI inside an active Windows implant

· 3 min read
Gemini 4 Carbon's cyber scores require a separate deployment gate

· 3 min read
METR's agent-monitor audit makes coverage accounting a security control
· 3 min read
Storm-0940's 2,900 sign-ins a minute make app-only identity the response boundary
· 3 min read
OpenCode's Console breach makes external API keys an enterprise incident boundary
· 3 min read
NIST's ten AI incident practices put agent containment ahead of recovery
· 2 min read
Open-weight model risk needs deployment controls, not an openness label

· 3 min read
Programmable DLP turns web gateways into software that must be audited
· 3 min read
Orca's 93% token reduction makes schema selection a security control
· 2 min read
Cloudflare's log transformers move data minimisation to the export boundary
· 2 min read
Cloudflare's new profiler exposes the hidden cost of production agents
· 2 min read
17,610 fake GitHub repositories make the README a security boundary
· 2 min read
State-linked data theft makes identity telemetry a national-security control
· 2 min read
Malware that talks to its analyst is a trust-boundary test
· 2 min read
Real-time phishing defeats codes when the session is the prize
· 2 min read
Anthropic's three cyber tiers turn model safeguards into an access-control system
· 2 min read
Agent security needs contextual policy, not another static permission screen

· 2 min read
Kubernetes operators are privileged identities. Audit the controller before trusting the workload.

· 2 min read
AI vulnerability reports need evidence gates before they reach human triage

· 2 min read
A self-hosted AI gateway is a privileged execution boundary, not just a model proxy

· 2 min read
474 live GitHub App keys make rotation an identity control, not a scanning task

· 2 min read
Gemini 4 Argon is gated by capability. Treat model access like privileged access.

· 2 min read
12% exploit success is enough to change patch policy. Open-weight cyber models compress defender time.

· 2 min read
Apple’s CoreGraphics zero-day makes OS branch inventory a security control, not an asset-management detail.

· 2 min read
One packet can stop telemetry. Patch TDengine before tuning the firewall.

· 2 min read
A watchdog inside the agent is still inside the trust boundary. Move enforcement out of band.

· 2 min read
Automated traffic passed humans in May. Rate-limit agents by cost, not identity.
· 2 min read
Eight speaker labels are not identity. Voice agents still need an authorisation check.

· 2 min read
Delegation is an authorisation boundary. Give every agent hand-off its own policy check.

· 2 min read
Sophistication stopped identifying the operator. Detection has to score the workflow.

· 2 min read
A DNS resolver became an internet gateway. Agent sandboxes need transitive egress controls.

· 2 min read
Seven minutes is enough to erase a cloud estate. Put recovery controls outside workload identity.

· 2 min read
A security widget without backend verification is decoration. Make the agent prove the control.

· 2 min read
A trusted agent can become the phisher. Put confirmation on every write action.

· 2 min read
MCP clients do not need cloud credentials. Put identity at the gateway.

· 2 min read
AI malware leaves cognitive artefacts. Hunt the plumbing before reverse engineering the payload.

· 2 min read
Prompt injection can be a defensive control. Keep the canary when the bomb stops working.

· 2 min read
Autonomous C2 changes the detection boundary. Watch the model call, not only the operator.

· 2 min read
Thousands of tenant agents do not need internet access. Make exfiltration fail at the network layer.

· 2 min read
Agentic ransomware changes the recovery target. Back up the model pipeline, not just the database.

· 2 min read
Plugin4Shell turns a pinned commit into an assertion. Verify the code that actually runs.
· 2 min read
Four coding agents crossed their sandboxes. The host trust graph is the boundary.

· 2 min read
A cyber agent crossed the test boundary. Prompts are not scope controls.

· 3 min read
72 hours to an internal repository changes the patching clock, not the vulnerability class

· 2 min read
Five browser agents fell to one extension. Extension policy is now an agent security control.

· 2 min read
3% false positives is not an AI scanner win. The threat model is doing the precision work.

· 2 min read
An encrypted credential vault is not a runtime boundary. Agent tools can still see plaintext.

· 2 min read
AI can find the flaw faster than the vendor can patch it. Compensating controls become a release requirement.

· 2 min read
Agent permissions are not team permissions. They need their own deployment boundary.
· 2 min read
The industry's safety warning is not a pause. It is a demand for deployer stop conditions.
· 2 min read
Click rate stopped measuring what phishing tests think it measures
· 2 min read
Your model registry runs on the same software CISA just flagged as under attack
· 2 min read
A million personalised fraud emails in three days breaks a defence industry's core assumption
· 2 min read
Your SOC's fastest-growing alert source is not an attacker. It is your own staff.
· 2 min read
1.8 million apps were already scanned for secrets. Assume yours was one of them.
· 2 min read
Account bans stop misuse of the model. They don't stop the system already built with it.

· 2 min read
200,000 exposed MCP servers is what happens when a protocol ships before its threat model
· 3 min read
Three labs shipped cyber models in one week. The capability is not the story.
· 2 min read
Tool poisoning works because the model cannot tell a description from an instruction
· 2 min read
A CVSS 10.0 in an agent framework is a different kind of vulnerability
· 2 min read
Prompt injection stopped being a content problem the moment it reached RCE
· 2 min read
Guardrails that run on a CPU change where you can put them
· 2 min read
The security question is not what your AI reads. It is what it can do.
· 2 min read
Six MCP incidents, one pattern: the credential outlived the task