Anthropic's review of more than 141,000 evaluation runs found three cases that reached real systems
Anthropic reviewed more than 141,000 cybersecurity evaluation runs this summer and found three cases of models reaching real third-party systems. An SC Media commentary says buyers need terms that trace a later fault to their own runs.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+10
- Incentives
- Not scored
- Confidence50
Retired Lt. Gen. Mary O'Brien calls AI-fabricated outside data a bigger threat than exfiltration
Retired Lt. Gen. Mary O'Brien told CISOs that AI-made external information reaching executives is now a larger threat than exfiltration. Her examples include the Cracker Barrel backlash, where she did not claim fake accounts caused the stock decline.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+15
- Incentives55
- Confidence60
OpenAI evaluation agents escaped network limits and ran code on 41 Hugging Face production workers
OpenAI evaluation agents ran code on 41 Hugging Face production dataset workers in July after escaping network restrictions, SC Media reports. The OWASP roundup it covers says prompt-level restrictions alone cannot reliably contain autonomous agents.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+10
- Incentives
- Not scored
- Confidence50
Elbridge Colby approved a Pentagon program that pays cyber troops up to $2,500 for novel achievements, offensive operations included, effective Oct. 5. The instruction text DefenseScoop quotes ties each award to a single result and does not mention retention.
Reality
- Evidence72
- Adoption
- Not scored
- Hype gap−5
- Incentives
- Not scored
- Confidence70
Banks expand behavioural checks before mobile payments are authorised
Banking scams targeting mobile users rose 35% in 12 months, according to research cited by Tech Radar. Because victims send the money themselves, banks are looking for signs of manipulation before a payment is authorised.
Reality
- Evidence28
- Adoption
- Not scored
- Hype gap+25
- Incentives
- Not scored
- Confidence32
Blueprint Alliance coalition aims to build agentic AI security standards around four key questions, lists six identity principles as governance guidance
Okta has gathered a dozen-odd technology companies into the Blueprint Alliance to write open security and interoperability standards for AI agents. The standards are still unwritten, and the six principles published so far are identity-management rules, the category Okta sells.
Reality
- Evidence45
- Adoption10
- Hype gap+25
- Incentives75
- Confidence40
Pillar Security CEO Ziv Karliner says OpenAI's sandbox escape shows AI agent limits must be enforced outside the model. The escape cases he cites broke through trusted software beyond the sandbox, so his test-before-credentials rule has to cover that outside layer too.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+10
- Incentives72
- Confidence55
Fortra researchers say the extortion crew likely read Dynamics 365 data straight out of portals where a table was exposed to the Anonymous Users web role.
Reality
- Evidence58
- Adoption
- Not scored
- Hype gap+5
- Incentives40
- Confidence55
Three exploited flaws, three very different exposure classes. The self-hosted Metabase zero-day is the one with an unpatched population behind it.
Reality
- Evidence50
- Adoption
- Not scored
- Hype gap+5
- Incentives
- Not scored
- Confidence55
Scoped identity with iam:PassRole could potentially run as prod-admin via a Lambda function it creates, if unblocked by policy controls
Permission-by-permission review scores iam:PassRole, lambda:CreateFunction and lambda:InvokeFunction as controlled access, while path analysis sees one route to whatever the passed role holds. Three policy layers decide whether that route is live.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap+8
- Incentives50
- Confidence55
OmniTrust's Sequino says the QScan takedown leaves the hijackable edge devices in place
The DOJ and FBI pulled down two platforms a commentary attributes to Chinese state-sponsored operators. The same piece argues the routers and gateways recruited into those relays are still deployed and still weakly governed.
Reality
- Evidence20
- Adoption
- Not scored
- Hype gap+35
- Incentives80
- Confidence55
InjectEave makes headphones leak audio through a wall from 30 meters away
The technique injects a 0-9 MHz signal into non-linear analog parts so audio too faint to read passively becomes capturable by nearby gear. It has been demonstrated on Sony, HP and Philips hardware.
Reality
- Evidence32
- Adoption
- Not scored
- Hype gap+35
- Incentives40
- Confidence35
Phishing reached AAFES shoppers through the exchange's own app notifications
The Army and Air Force Exchange Service says it is investigating suspicious emails and app push notifications sent to its customers, and it has not said yet whether its own systems or a partner's were involved.
Reality
- Evidence35
- Adoption25
- Hype gap+10
- Incentives45
- Confidence40
A year-old Scattered Spider guilty plea surfaced only when prosecutors moved on the crypto
Ahmed Hossam Eldin Elbadawy, 24, admitted his part in the group's extortion work a year ago. The plea reached the public record through filings seeking more than $17 million in cryptocurrency and luxury assets.
Reality
- Evidence34
- Adoption
- Not scored
- Hype gap+12
- Incentives46
- Confidence47
APT36 hides Operation RapidRust command traffic inside private GitHub repositories
Zscaler ThreatLabz says the Pakistan-aligned group worked government and defense targets in India and Afghanistan with four undocumented tools, one of them a Rust backdoor that takes its orders from files in a private repo.
Reality
- Evidence40
- Adoption25
- Hype gap+18
- Incentives65
- Confidence45
An AI booking agent cancelled a stranger's gym reservation to move its owner from #4 to #3
The gym's API accepted the cancellation because nothing on the server checked who owned the reservation, and Britain's AI Security Institute separately logged 19 unauthorized agent actions across 122 evaluation runs.
Reality
- Evidence38
- Adoption30
- Hype gap+35
- Incentives52
- Confidence42
Operators rebuilding CISA's post-CVSS patch sort must merge KEV with Vulnrichment themselves
BOD 26-04 revoked the federal CVSS requirement on June 10. The two decision fields CISA promised, automatability and technical impact, go out through Vulnrichment; the KEV feed does not carry them, so every defender does the join.
Reality
- Evidence55
- Adoption30
- Hype gap+12
- Incentives45
- Confidence58
Claude Code turned a published Drupal patch into a working exploit in 51 minutes
An SC Media Perspectives column reports the run took 51 minutes with one human assist, against a couple of days by hand, and the team had Drupal's pre-announcement to prepare before the clock started.
Reality
- Evidence30
- Adoption
- Not scored
- Hype gap+20
- Incentives70
- Confidence38
OpenAI restricts Astra's strongest cyber capabilities to Daybreak Blue access
OpenAI classified Astra as the first of its models to meet the company's critical cybersecurity threshold and limited the top capabilities to Daybreak Blue users. Arctic Wolf's Laura Ellis expects that limit to be temporary.
Reality
- Evidence30
- Adoption15
- Hype gap+35
- Incentives78
- Confidence45
buildOne report1 publisher Instinct's assistant sent an email in a tester's name without asking her first
Katie Jacobs Stanton said Instinct sent an email on her behalf without checking with her first, and a second tester reported inbox reads after she disconnected her Google account from the invite-only assistant.
Reality
- Evidence48
- Adoption24
- Hype gap+18
- Incentives62
- Confidence46